Text-assisted traffic prediction system and method

Abstract

The application discloses a text auxiliary-based traffic prediction system and method, and the system comprises a text reasoning and processing module, a time series feature extraction module, a multi-scale modal fusion module and a self-adaptive prediction output module which are connected in sequence, and each module cooperates to complete a network traffic prediction task. Through deep reasoning modeling of external text, multi-scale decomposition and fusion of time series features and a robust optimization mechanism, the application can realize the coordinated improvement of technical indexes such as network traffic prediction accuracy, efficiency and robustness.

Description

Technical Field

[0001] This invention relates to a text-assisted traffic prediction system and method, belonging to the field of information technology, and particularly to the field of intelligent network information technology. Background Technology

[0002] Network traffic prediction is one of the main tasks of intelligent operation and maintenance of communication networks. This technology is widely used in important scenarios such as network bandwidth management and network resource scheduling by operators. Improving prediction accuracy helps optimize network resource utilization and the stability of network service quality. With the continuous development of Internet technology, various network devices (routers, switches, edge gateways) continuously generate a large amount of network traffic data, including real-time bandwidth usage, packet transmission rate, and website traffic, providing fundamental data support for accurate network traffic prediction. Furthermore, the rapid advancement of multimodal technology has promoted the deep integration of text information with network traffic prediction tasks: external texts such as network operation and maintenance notices, bandwidth adjustment announcements, and network equipment maintenance plans contain key driving factors that directly affect network traffic trend fluctuations. Effectively integrating this information into prediction models has become an important research direction for improving network traffic prediction performance.

[0003] With breakthroughs in cross-modal understanding and semantic reasoning capabilities of large language models, related research has begun to explore using these models to extract textual value and assist in web traffic prediction. However, most existing solutions remain at the level of simply using large language models for text encoding or information filtering, failing to fully leverage the complementary roles of text and web traffic. Some solutions directly use large language models as the core component of the prediction model, attempting to enhance feature representation but ignoring the inherent limitations of large language models in processing time-series data. This not only fails to significantly improve prediction performance but also leads to a sharp increase in computational costs, making it difficult to meet the efficiency requirements of practical applications. Other solutions merely use large language models to distill and summarize the text, attempting to filter redundant information. However, this superficial processing method still cannot fully extract the deep semantics closely related to web traffic trends in the text, wasting a large amount of valuable information.

[0004] Existing network traffic prediction schemes incorporating external text still face numerous unresolved technical challenges. First, there is often a lack of direct and explicit semantic connection between external text and network traffic. The text content primarily describes the events themselves in natural language, without directly pointing to the changing trends of network traffic. This makes it difficult for models to accurately identify information crucial for prediction, especially when dealing with new events not encountered during the training phase, making it impossible to effectively determine the direction and extent of the text's influence on network traffic. Second, network traffic itself exhibits significant multi-scale characteristics. Signals in different frequency domains correspond to different dimensions of information, such as instantaneous fluctuations, short-term changes, and long-term trends. Different external events have varying impacts on these different scale characteristics. Existing modality fusion methods fail to specifically utilize this inherent characteristic of network traffic, leading to entanglement of temporal information across different frequency domains, insufficient matching accuracy between text information and network traffic features, and significantly reduced fusion effectiveness. Furthermore, existing solutions struggle to balance prediction accuracy and computational efficiency when incorporating textual information. Often, the pursuit of performance gains comes at the cost of excessive computational complexity, failing to meet the high-efficiency requirements of traditional network traffic prediction models. Simultaneously, textual data inevitably contains noise and may even exhibit semantic biases. Existing models lack effective mechanisms to adapt to this noise, leading to decreased prediction robustness and difficulty in handling complex real-world application scenarios. These intertwined problems collectively constrain further improvements in network traffic prediction performance. Addressing these issues has become a pressing technical challenge in the field of network traffic prediction. Summary of the Invention

[0005] In view of this, the purpose of this invention is to invent a system and method that, through deep reasoning modeling of external text, multi-scale decomposition and fusion of temporal features, and robustness optimization mechanism, achieves a synergistic improvement in technical indicators such as network traffic prediction accuracy, efficiency, and robustness.

[0006] To achieve the above objectives, this invention proposes a text-assisted network traffic prediction system. The system comprises a text reasoning and processing module, a network traffic feature extraction module, a multi-scale modality fusion module, and an adaptive prediction output module, connected sequentially. These modules work together to complete the network traffic prediction task, as detailed below:

[0007] The text reasoning and processing module consists of a large language model reasoning unit, a self-checking filtering unit, and a text encoding unit connected in series. Its core function is to perform deep semantic reasoning transformation, illusion and redundant noise filtering, and high-dimensional feature encoding on various external texts such as network operation and maintenance notices, bandwidth adjustment announcements, and large-scale event announcements. It transforms unstructured natural language text into high-quality text embeddings that are directly related to the future trend of network traffic and have clear semantics, providing a direct semantic association between external text and network traffic for the subsequent modality fusion process.

[0008] Network traffic feature extraction module: It adopts a plug-and-play architecture design and is compatible with mainstream network traffic prediction models as the backbone model. It consists of a time series preprocessing unit and a multi-scale frequency decomposition unit connected in series. It can comprehensively capture the inherent time dependence, periodic fluctuations, seasonal variation patterns and multi-scale frequency features of network traffic. While generating preliminary prediction results, it simultaneously outputs multi-scale time series tensors corresponding to high, medium and low frequency domains, laying the foundation for hierarchical modal fusion.

[0009] Multi-scale modality fusion module: Composed of multi-scale convolutional units and cross-attention fusion units connected in series, it achieves accurate matching and deep fusion of text embedding and temporal features at different frequency domain levels, effectively avoiding the entanglement of temporal information in different frequency domains, giving full play to the complementary role of text semantic information and temporal features at different scales, and improving the effectiveness and pertinence of modality fusion.

[0010] The adaptive prediction output module consists of an adaptive loss calculation unit and a projection output unit connected in series. It dynamically adjusts the prediction bias through the adaptive loss function, effectively reducing the interference of text noise and illusion information on the prediction results. At the same time, it completes the feature dimension mapping through the projection output unit, and finally outputs accurate network traffic prediction results that meet the actual application requirements within a set future time period, balancing prediction accuracy and robustness.

[0011] The specific working process of the text reasoning and processing module includes the following tasks:

[0012] (1) The large language model reasoning unit adopts a large language model with cross-modal semantic reasoning capabilities. It takes historical network traffic data, domain background information and external text as input, and generates a unified format prediction text through chain thinking CoT. This text directly describes the direction, magnitude and key time window of the impact of external events on the future trend of network traffic. The reasoning process is automated through system prompts. Only the background information needs to be fine-tuned according to the changes in the dataset, without a lot of manual intervention.

[0013] (2) The self-checking filtering unit realizes text illusion detection based on the natural language inference (NLI) mechanism. It takes the predicted text generated by the large language model as the hypothesis and samples key paragraphs from the external text as the premise. It obtains the NLI score by calculating the contradiction probability between the two. When the score exceeds the set threshold, it is judged as illusory text and directly removed to ensure the reliability of text information.

[0014] (3) The text encoding unit uses a pre-trained BERT model to generate a text tensor for the effective predicted text encoding, freezes the model parameters, and fills the zero vector at time steps without corresponding external text.

[0015] The formula for calculating the probability of contradiction is:

[0016]

[0017] in, and These are the log probabilities for the categories "implication" and "contradiction," respectively. To predict text, This is a sampled section;

[0018] The formula for calculating the NLI score is as follows:

[0019]

[0020] Where N is the number of sampled segments. The sequence number of the predicted text generated by the large language model. .

[0021] The specific working process of the time series feature extraction module includes the following tasks:

[0022] (1) The time series preprocessing unit standardizes the historical network traffic, captures local semantics and long-range dependencies through the backbone model, and generates initial prediction results;

[0023] (2) The multi-scale frequency decomposition unit decomposes the initial prediction result into features in different frequency domains through Discrete Fourier Transform (DFT). After being partitioned by a frequency selection function, it is converted into a multi-scale time series tensor in the time domain through Inverse Discrete Fourier Transform (IDFT). The calculation formula of the DFT is as follows:

[0024]

[0025] in Let be the initial prediction result in the time domain, and len be the total length of the network traffic in the initial prediction result. The time step takes a value from 0 to len−1. For frequency values, The imaginary unit, This represents the initial prediction result in the frequency domain.

[0026] The frequency selection function is defined as:

[0027]

[0028] in For the frequency selection function of the i-th frequency band, Let be the lower limit frequency of the i-th frequency band. Let i be the upper limit frequency of the i-th frequency band. The value to be determined is the frequency value.

[0029] The specific working process of the multi-scale modality fusion module includes the following tasks:

[0030] (1) Multi-scale convolutional units employ convolutional operations with differentiated dilation rates for temporal tensors of different frequency levels. High-frequency temporal tensors correspond to the rapid fluctuation characteristics of network traffic, and convolutional kernels with small dilation rates are used to accurately capture instantaneous change details. Low-frequency temporal tensors correspond to long-term stable trends, and convolutional kernels with large dilation rates are used to fully extract global trend information. The dilation rate of medium-frequency temporal tensors is between the two. Through differentiated convolution processing, the expressive power of temporal features in different frequency domains is enhanced, providing more targeted feature inputs for subsequent fusion.

[0031] (2) The cross-attention fusion unit first performs linear projection transformation on the convolutional temporal tensors and text tensors at each scale. The temporal tensors at each scale are projected to generate query vectors Q, and the text tensors are projected to generate unified key vectors K and value vectors V. Then, attention weights are obtained by calculating the similarity between the query vectors and key vectors. These weights reflect the correlation strength between text semantic information and temporal features at each scale. Finally, the value vectors are weighted and aggregated based on the attention weights to obtain multi-scale fusion features that fuse text semantics and corresponding temporal features, thereby realizing the deep association between text and temporal data.

[0032] The formula for calculating the convolution operation is:

[0033] in Let be the weight value at the s-th position in the convolution kernel, where s is the index of the convolution kernel. Let be the value of the original time series tensor of the i-th frequency band at time step t. Let be the expansion rate of the i-th frequency band. This refers to the mapping time step after the convolution operation; The formula for calculating the fused features in the above steps is:

[0034] ,

[0035] in, Let be the fused feature tensor obtained by fusing the temporal features and textual features of the i-th frequency band. For query vector dimensions, The query vector generated by linear projection of the time series tensor of the i-th frequency band. It is the transpose of the key vector generated by linear projection of the text tensor. It is a value vector generated by linear projection of the text tensor.

[0036] The specific working process of the adaptive prediction output module includes the following tasks:

[0037] (1) The adaptive loss calculation unit calculates the initial prediction loss and the fusion prediction loss, and dynamically adjusts the loss value through the adaptive loss function. In the later stage of training, the fusion prediction loss is smoothly transitioned to the adaptive loss.

[0038] (2) The projection output unit uses a 1×1 standard convolutional layer to map the multi-scale fused feature tensor to the target dimension, and outputs the predicted network traffic value within a set future time period. The formula for calculating the adaptive loss function is:

[0039]

[0040] in, The output loss value of the adaptive loss function. For the initial predicted loss, To fuse the prediction loss, Scale is the scaling factor for loss adjustment, used to adjust the magnitude of the loss function value, and Offset is the offset factor for loss adjustment, used to fine-tune the input baseline value of the loss function.

[0041] This invention also proposes a text-assisted network traffic prediction method, which includes the following steps:

[0042] (1) Text processing and embedding generation: Historical network traffic data, corresponding domain background information and various external texts are input into the inference unit of the large language model. The predicted text conforming to a unified format is generated by the chain thinking CoT. The NLI score of each predicted text is calculated by the self-checking filter unit, and the phantom text with the score exceeding the threshold is removed. The filtered effective predicted text is input into the pre-trained BERT encoder, the model parameters are frozen for encoding processing, and a high-dimensional text tensor is generated. The zero vector is filled in the time step where there is no corresponding external text.

[0043] (2) Temporal feature extraction and decomposition: The input historical network traffic is standardized to eliminate the difference in dimensions; the local semantics and long-range dependence of the network traffic are captured by the selected backbone model to generate the initial prediction results; the initial prediction results are converted to the frequency domain by using Discrete Fourier Transform (DFT), and three non-overlapping sub-bands of high, medium and low frequency are divided by the frequency selection function. Then, the signals of each sub-band are converted back to the time domain by Inverse Discrete Fourier Transform (IDFT) to obtain the multi-scale temporal tensor.

[0044] (3) Multi-scale modal fusion: Perform convolution operations with different dilation rates on the temporal tensors of each scale to enhance the expression of features in different frequency domains; perform linear projection of the convolved temporal tensors and the text tensors to obtain query vectors, key vectors and value vectors; calculate attention weights and aggregate value vectors to achieve deep fusion of text semantics and temporal features of each scale, and obtain multi-scale fusion features.

[0045] (4) Adaptive prediction output: Calculate the initial prediction loss and the fusion prediction loss respectively, and dynamically adjust the loss value through the adaptive loss function to achieve a smooth transition of loss during training; input the multi-scale fusion features into the standard convolutional layer for dimension mapping, and output the network traffic prediction value within the future set time period.

[0046] The beneficial effects of this invention are as follows: It utilizes the semantic reasoning capabilities of a large language model to transform external text into predicted text directly related to network traffic, solving the core problem of the disconnect between text and temporal semantics. Simultaneously, it filters illusory text through a self-checking mechanism, improving the reliability of text information. Based on a multi-scale modal fusion design in the frequency domain, it fully adapts to the inherent multi-scale characteristics of network traffic, achieving accurate matching of text and temporal information across different frequency domains, avoiding cross-scale information entanglement, and significantly improving the fusion effect. It employs an adaptive loss function to dynamically address text noise, is compatible with existing efficient temporal models, controls computational complexity while ensuring prediction accuracy, and balances robustness and efficiency. It exhibits excellent adaptability to both stationary and non-stationary network traffic temporal data, making it highly valuable for practical applications. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the modules of the text-assisted traffic prediction system proposed in this invention;

[0049] Figure 2 This is a schematic diagram illustrating the working process of the text reasoning and processing module in this embodiment of the invention;

[0050] Figure 3 This is a schematic diagram of the large language model reasoning unit of the text reasoning and processing module in this embodiment of the invention;

[0051] Figure 4 This is a schematic diagram illustrating the working process of the adaptive prediction output module in an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram illustrating the working process of the network traffic feature extraction module and the multi-scale modality fusion module in this embodiment of the invention.

[0053] Figure 6 This is a flowchart illustrating the text-assisted network traffic prediction method proposed in this invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0056] See Figure 1This invention introduces a text-assisted network traffic prediction system. The system comprises a text reasoning and processing module, a network traffic feature extraction module, a multi-scale modality fusion module, and an adaptive prediction output module, which are connected in sequence. Each module works together to complete the network traffic prediction task.

[0057] The text reasoning and processing module consists of a large language model reasoning unit, a self-checking filtering unit, and a text encoding unit connected in series. Its core function is to perform deep semantic reasoning transformation, illusion and redundant noise filtering, and high-dimensional feature encoding on various external texts such as network operation and maintenance notices, bandwidth adjustment announcements, and large-scale event announcements. It transforms unstructured natural language text into high-quality text embeddings that are directly related to the future trend of network traffic and have clear semantics, providing a direct semantic association between external text and network traffic for the subsequent modality fusion process.

[0058] Network traffic feature extraction module: It adopts a plug-and-play architecture design and is compatible with mainstream network traffic prediction models as the backbone model. It consists of a time series preprocessing unit and a multi-scale frequency decomposition unit connected in series. It can comprehensively capture the inherent time dependence, periodic fluctuations, seasonal variation patterns and multi-scale frequency features of network traffic. While generating preliminary prediction results, it simultaneously outputs multi-scale time series tensors corresponding to high, medium and low frequency domains, laying the foundation for hierarchical modal fusion.

[0059] Multi-scale modality fusion module: Composed of multi-scale convolutional units and cross-attention fusion units connected in series, it achieves accurate matching and deep fusion of text embedding and temporal features at different frequency domain levels, effectively avoiding the entanglement of temporal information in different frequency domains, giving full play to the complementary role of text semantic information and temporal features at different scales, and improving the effectiveness and pertinence of modality fusion.

[0060] The adaptive prediction output module consists of an adaptive loss calculation unit and a projection output unit connected in series. It dynamically adjusts the prediction bias through the adaptive loss function, effectively reducing the interference of text noise and illusion information on the prediction results. At the same time, it completes the feature dimension mapping through the projection output unit, and finally outputs accurate network traffic prediction results that meet the actual application requirements within a set future time period, balancing prediction accuracy and robustness.

[0061] See Figure 2 The specific working process of the text reasoning and processing module includes the following tasks:

[0062] (1) The large language model reasoning unit adopts a large language model with cross-modal semantic reasoning capabilities. It takes historical network traffic data, domain background information and external text as input, and generates a unified format prediction text through chain thinking CoT. This text directly describes the direction, magnitude and key time window of the impact of external events on the future trend of network traffic. The reasoning process is automated through system prompts. Only the background information needs to be fine-tuned according to the changes in the dataset, without a lot of manual intervention.

[0063] (2) The self-checking filtering unit realizes text illusion detection based on the natural language inference (NLI) mechanism. It takes the predicted text generated by the large language model as the hypothesis and samples key paragraphs from the external text as the premise. It obtains the NLI score by calculating the contradiction probability between the two. When the score exceeds the set threshold, it is judged as illusory text and directly removed to ensure the reliability of text information.

[0064] (3) The text encoding unit uses a pre-trained BERT model to generate a text tensor for the effective predicted text encoding, freezes the model parameters, and fills the zero vector at time steps without corresponding external text.

[0065] The formula for calculating the probability of contradiction is:

[0066]

[0067] in, and These are the log probabilities for the categories "implication" and "contradiction," respectively. To predict text, This is a sampled section;

[0068] The formula for calculating the NLI score is as follows:

[0069]

[0070] Where N is the number of sampled segments. The sequence number of the predicted text generated by the large language model. .

[0071] In this embodiment, the self-checking filtering unit includes sampling three key paragraphs from the external text as a premise, generating a predicted text as a hypothesis, and using a Natural Language Inference (NLI) model to calculate the contradiction probability. The NLI score threshold is 0.9; if the score exceeds the threshold, it is determined to be illusory text and is discarded. Text encoding: A pre-trained BERT-based model is selected as the encoder, and the Transformer layer parameters are frozen to reduce the computational load; the effective predicted text is segmented into words, with a maximum word count. If there is no corresponding external text, fill the time step with zero vectors and output the text tensor.

[0072] See Figure 3 The working process of the large language model reasoning unit of the text reasoning and processing module includes the following tasks:

[0073] (1) Input integration: Historical network traffic data, corresponding domain background information and various external texts are integrated as input to the reasoning unit of the large language model. The domain background information includes the inherent characteristics of the time series data in the domain and the typical influence of external events on the time series data. The external texts cover various text information related to network traffic changes, such as network maintenance notices, network attack alarms, and announcements of large-scale events.

[0074] (2) Chain reasoning guidance: Under the guidance of the preset system prompts, the reasoning unit of the large language model uses the chain thinking (CoT) method to complete the reasoning step by step. First, it identifies the key events and core attributes in the external text, and then combines the historical network traffic change pattern and domain background knowledge to analyze the potential impact mechanism of the key event on the future trend of network traffic. Finally, it quantifies the direction, magnitude and key time window of the event's impact.

[0075] (3) Structured text generation: After the reasoning is completed, a prediction text conforming to a unified preset format is automatically generated. This text clearly and explicitly describes the specific impact of external events on the future trend of network traffic, ensuring that the text content is directly related to the network traffic prediction task, and providing clear semantic association for subsequent text encoding and modal fusion.

[0076] (4) Automated adaptation: The entire predictive text generation process is fully automated through system prompts, without the need for a lot of manual intervention. Only the domain background information needs to be fine-tuned according to the characteristics of different domain datasets to adapt to the prediction needs in different scenarios, ensuring the scalability and ease of use of the method.

[0077] In this embodiment, the predicted text generation uses the Qwen-MAX large language model as the core model. The system prompts consist of historical features of network traffic time series data and domain-specific background. The model input is historical network traffic data combined with domain-adapted external text, such as network maintenance notices, network attack alarms, and announcements of large-scale events. The model finally outputs predicted text in a unified format, such as "Based on historical data, it is expected that website access traffic will increase by 35% from 19:00 to 22:00 tonight, and reach its peak at 8 pm."

[0078] See Figure 4 The specific working process of the adaptive prediction output module includes the following:

[0079] (1) The time series preprocessing unit standardizes the historical network traffic, captures local semantics and long-range dependencies through the backbone model, and generates initial prediction results;

[0080] (2) The multi-scale frequency decomposition unit decomposes the initial prediction result into different frequency domain features through the Discrete Fourier Transform (DFT). After being divided by the frequency selection function, it is converted into a multi-scale time series tensor in the time domain through the Inverse Discrete Fourier Transform (IDFT). Then, it is fused with the text vector to generate the initial prediction result.

[0081] (3) The adaptive loss calculation unit calculates the initial prediction loss and the fusion prediction loss, and dynamically adjusts the loss value through the adaptive loss function. In the later stage of training, the fusion prediction loss is smoothly transitioned to the adaptive loss.

[0082] (4) The projection output unit uses a 1×1 standard convolutional layer to map the multi-scale fused feature tensor to the target dimension, and outputs the predicted network traffic value within a set future time period. The formula for calculating the adaptive loss function is:

[0083]

[0084] in, The output loss value of the adaptive loss function. For the initial predicted loss, To fuse the prediction loss, Scale is the scaling factor for loss adjustment, used to adjust the magnitude of the loss function value, and Offset is the offset factor for loss adjustment, used to fine-tune the input baseline value of the loss function.

[0085] In this embodiment, the adaptive loss calculation method involves calculating the initial prediction loss. and fusion prediction loss ( (For the true value); set the adjustment parameters Scale=0.5, Offset=0.1, and the adaptive loss function is...

[0086] Sigmoid In the later stages of training, from Smooth transition to Projection output: using Standard convolutional layers map the fused feature tensors to the target dimension and output the final prediction result. , i.e., the future The time series prediction value at each time step.

[0087] See Figure 5 The specific working process of the time series feature extraction module and the multi-scale modality fusion module includes the following:

[0088] (1) Multi-scale convolutional units employ convolutional operations with differentiated dilation rates for temporal tensors of different frequency levels. High-frequency temporal tensors correspond to the rapid fluctuations in network traffic, and convolutional kernels with small dilation rates are used to accurately capture instantaneous change details. Low-frequency temporal tensors correspond to long-term stable trends, and convolutional kernels with large dilation rates are used to fully extract global trend information. The dilation rate of medium-frequency temporal tensors is between the two. Through differentiated convolution processing, the expressive power of temporal features in different frequency domains is enhanced, providing more targeted feature inputs for subsequent fusion.

[0089] The formula for calculating the DFT is as follows:

[0090]

[0091] in Let be the initial prediction result in the time domain, and len be the total length of the network traffic in the initial prediction result. The time step takes a value from 0 to len−1. For frequency values, The imaginary unit, This represents the initial prediction result in the frequency domain.

[0092] The frequency selection function is defined as:

[0093]

[0094] in For the frequency selection function of the i-th frequency band, Let be the lower limit frequency of the i-th frequency band. Let i be the upper limit frequency of the i-th frequency band. The value to be determined is the frequency value.

[0095] (2) Multi-scale convolutional units employ convolutional operations with differentiated dilation rates for temporal tensors of different frequency levels. High-frequency temporal tensors correspond to the rapid fluctuation characteristics of network traffic, and convolutional kernels with small dilation rates are used to accurately capture instantaneous change details. Low-frequency temporal tensors correspond to long-term stable trends, and convolutional kernels with large dilation rates are used to fully extract global trend information. The dilation rate of medium-frequency temporal tensors is between the two. Through differentiated convolution processing, the expressive power of temporal features in different frequency domains is enhanced, providing more targeted feature inputs for subsequent fusion.

[0096] (3) The cross-attention fusion unit first performs linear projection transformation on the convolutional temporal tensors and text tensors at each scale. The temporal tensors at each scale are projected to generate query vectors Q, and the text tensors are projected to generate unified key vectors K and value vectors V. Then, attention weights are obtained by calculating the similarity between the query vectors and key vectors. These weights reflect the correlation strength between text semantic information and temporal features at each scale. Finally, the value vectors are weighted and aggregated based on the attention weights to obtain multi-scale fusion features that fuse text semantics and corresponding temporal features, thereby realizing the deep association between text and temporal data.

[0097] The formula for calculating the convolution operation is:

[0098]

[0099] in Let be the weight value at the s-th position in the convolution kernel, where s is the index of the convolution kernel. Let be the value of the original time series tensor of the i-th frequency band at time step t. Let be the expansion rate of the i-th frequency band. This represents the mapping time step after the convolution operation.

[0100] The formula for calculating the fused features in the above steps is:

[0101] ,

[0102] in, Let be the fused feature tensor obtained by fusing the temporal features and textual features of the i-th frequency band. For query vector dimensions, The query vector generated by linear projection of the time series tensor of the i-th frequency band. It is the transpose of the key vector generated by linear projection of the text tensor. It is a value vector generated by linear projection of the text tensor.

[0103] In this embodiment, historical network traffic data is first analyzed. Standardization was performed (mean 0, variance 1); PatchTST was selected as the backbone model, network traffic was divided into several patches, local semantics and long-range dependency features were extracted, and initial prediction results were generated. Then, multi-scale frequency decomposition is performed on the initial prediction results: Performing a Discrete Fourier Transform (DFT) yields the frequency domain signal. Three non-overlapping frequency bands are defined (high frequency corresponds to instantaneous fluctuations, mid frequency corresponds to short-term changes, and low frequency corresponds to long-term trends). The signals of each frequency band are converted back to the time domain using the inverse discrete Fourier transform (IDFT) to obtain multi-scale temporal tensors. In the subsequent fusion stage, a multi-scale convolutional unit first applies convolutional operations with differentiated dilation rates to the temporal tensors at different frequency levels. High-frequency tensors use convolutional kernels with small dilation rates to accurately capture instantaneous change details, while low-frequency tensors use convolutional kernels with large dilation rates to fully extract global trend information. The dilation rate of the mid-frequency tensor is between the two, enhancing the expressive power of temporal features in different frequency domains. Then, a cross-attention fusion unit performs linear projection on the convolved temporal tensors at each scale and the text tensor, transforming the temporal tensors at each scale into query vectors and the text tensors into unified key and value vectors. Attention weights are obtained by calculating the similarity between the query vector and the key vector, and then the value vectors are weighted and aggregated based on these weights. Finally, deep fusion of text semantics and temporal features at each scale is achieved, resulting in multi-scale fused features.

[0104] See Figure 6 This invention introduces a text-assisted network traffic prediction method, which includes the following steps:

[0105] (1) Text processing and embedding generation: Historical network traffic data, corresponding domain background information and various external texts are input into the inference unit of the large language model. The predicted text conforming to a unified format is generated by the chain thinking CoT. The NLI score of each predicted text is calculated by the self-checking filter unit, and the phantom text with the score exceeding the threshold is removed. The filtered effective predicted text is input into the pre-trained BERT encoder, the model parameters are frozen for encoding processing, and a high-dimensional text tensor is generated. The zero vector is filled in the time step where there is no corresponding external text.

[0106] (2) Temporal feature extraction and decomposition: The input historical network traffic is standardized to eliminate the difference in dimensions; the local semantics and long-range dependence of the network traffic are captured by the selected backbone model to generate the initial prediction results; the initial prediction results are converted to the frequency domain by using Discrete Fourier Transform (DFT), and three non-overlapping sub-bands of high, medium and low frequency are divided by the frequency selection function. Then, the signals of each sub-band are converted back to the time domain by Inverse Discrete Fourier Transform (IDFT) to obtain the multi-scale temporal tensor.

[0107] (3) Multi-scale modal fusion: Perform convolution operations with different dilation rates on the temporal tensors of each scale to enhance the expression of features in different frequency domains; perform linear projection of the convolved temporal tensors and the text tensors to obtain query vectors, key vectors and value vectors; calculate attention weights and aggregate value vectors to achieve deep fusion of text semantics and temporal features of each scale, and obtain multi-scale fusion features.

[0108] (4) Adaptive prediction output: Calculate the initial prediction loss and the fusion prediction loss respectively, and dynamically adjust the loss value through the adaptive loss function to achieve a smooth transition of loss during training; input the multi-scale fusion features into the standard convolutional layer for dimension mapping, and output the network traffic prediction value within the future set time period.

[0109] The inventors conducted extensive experiments on the proposed system and method. The experimental results are as follows: The inventors selected two publicly available time-series datasets in the field of network traffic to verify the generalization ability of the system. Specific information is shown in Table 1:

[0110] Table 1

[0111] Dataset Name Time Granularity Data Duration Sample Number Network Traffic 15 minutes 6 months 12344 Web 1 day 6 years 2167

[0112] Mean squared error (MSE) and mean absolute error (MAE) are used as the core evaluation indicators. The smaller the value, the higher the prediction accuracy. The formulas are defined as follows:

[0113]

[0114]

[0115] in, For the first The time step, the first Predicted values ​​for each dimension To correspond to the actual value, To predict the window length, For time-series dimensions, see Table 2 for experimental results on each dataset.

[0116] Table 2

[0117] Metrics Invention PatchTST Time-LLM GPT4TS Time-Mixer DLinear TimesNet Time-MMD Network Traffic-MSE 0.146 0.191 0.209 0.238 0.196 0.206 0.232 0.270 Network Traffic-MAE 0.251 0.269 0.263 0.315 0.282 0.282 0.279 0.349 Web-MSE 0.171 0.244 0.292 0.342 0.260 0.679 0.314 0.256 Web-MAE 0.311 0.345 0.393 0.445 0.349 0.649 0.401 0.355

[0118] The system of this invention achieved optimal performance across all metrics on all datasets, verifying that the text semantic enhancement brought about by large language model inference and the full utilization of temporal features by multi-scale modality fusion effectively improve prediction accuracy. Furthermore, to verify the necessity of the core module, three ablation experiments were designed, and the experimental results are shown in Table 3.

[0119] Table 3

[0120] Experimental Groups Network Traffic (MSE / MAE) Web (MSE / MAE) Complete System 0.129 / 0.238 0.171 / 0.311 w / o Inference Module 0.155 / 0.242 0.208 / 0.337 w / o Multi-scale Fusion 0.135 / 0.244 0.195 / 0.331 w / o Adaptive Loss 0.141 / 0.240 0.204 / 0.328

[0121] In the experimental groups in Table 3, removing the inference module (without inference module): directly fusing the original external text with the time-series data led to a significant decrease in prediction accuracy for each dataset, indicating a semantic disconnect between the original text and the time-series data. The predicted text generated by the large language model inference is key to improving performance. Removing multi-scale fusion (without multi-scale fusion): performing modal fusion only at a single scale cannot adapt to the multi-frequency features of the time-series data, resulting in an increase in MSE for each dataset, verifying that multi-scale fusion can accurately match the complementary information between text and time series. Removing adaptive loss (without adaptive loss): using a fixed loss function cannot cope with label errors caused by text noise and illusions, leading to a decrease in model robustness, proving that the adaptive loss function can effectively mitigate noise interference.

[0122] Experimental results show that the system and method proposed in this invention are effective and feasible, and can effectively improve the accuracy of network traffic prediction.

Claims

1. A text-assisted traffic prediction system, characterized in that: The system comprises a text reasoning and processing module, a time series feature extraction module, a multi-scale modality fusion module, and an adaptive prediction output module, which are connected in sequence. These modules work together to complete the network traffic prediction task, as detailed below: The text reasoning and processing module consists of a large language model reasoning unit, a self-checking filtering unit, and a text encoding unit connected in series. Its core function is to perform deep semantic reasoning transformation, illusion and redundant noise filtering, and high-dimensional feature encoding on various external texts such as network operation and maintenance notices, bandwidth adjustment announcements, and large-scale event announcements. It transforms unstructured natural language text into high-quality text embeddings that are directly related to the future trend of network traffic and have clear semantics, providing a direct semantic association between external text and network traffic for the subsequent modality fusion process. The time series feature extraction module adopts a plug-and-play architecture design, which is compatible with mainstream network traffic prediction models as the backbone model. It consists of a time series preprocessing unit and a multi-scale frequency decomposition unit connected in series. It can comprehensively capture the inherent time dependence, periodic fluctuations, seasonal variation patterns and multi-scale frequency features of network traffic. While generating preliminary prediction results, it simultaneously outputs multi-scale time series tensors corresponding to high, medium and low frequency domains, laying the foundation for hierarchical modal fusion. Multi-scale modality fusion module: Composed of multi-scale convolutional units and cross-attention fusion units connected in series, it achieves accurate matching and deep fusion of text embedding and temporal features at different frequency domain levels, effectively avoiding the entanglement of temporal information in different frequency domains, giving full play to the complementary role of text semantic information and temporal features at different scales, and improving the effectiveness and pertinence of modality fusion. The adaptive prediction output module consists of an adaptive loss calculation unit and a projection output unit connected in series. It dynamically adjusts the prediction bias through the adaptive loss function, effectively reducing the interference of text noise and illusion information on the prediction results. At the same time, it completes the feature dimension mapping through the projection output unit, and finally outputs accurate network traffic prediction results that meet the actual application requirements within a set future time period, balancing prediction accuracy and robustness.

2. The text-assisted traffic prediction system according to claim 1, characterized in that: The specific working process of the text reasoning and processing module includes the following tasks: (1) The large language model reasoning unit adopts a large language model with cross-modal semantic reasoning capabilities. It takes historical network traffic data, domain background information and external text as input, and generates a unified format prediction text through chain thinking CoT. This text directly describes the direction, magnitude and key time window of the impact of external events on the future trend of network traffic. The reasoning process is automated through system prompts. Only the background information needs to be fine-tuned according to the changes in the dataset, without a lot of manual intervention. (2) The self-checking filtering unit realizes text illusion detection based on the natural language inference (NLI) mechanism. It takes the predicted text generated by the large language model as the hypothesis and samples key paragraphs from the external text as the premise. It obtains the NLI score by calculating the contradiction probability between the two. When the score exceeds the set threshold, it is judged as illusory text and directly removed to ensure the reliability of text information. (3) The text encoding unit uses a pre-trained BERT model to generate a text tensor for the effective predicted text encoding, freezes the model parameters, and fills the zero vector at time steps without corresponding external text. The formula for calculating the probability of contradiction is: in, and These are the log probabilities for the categories "implication" and "contradiction," respectively. To predict text, This is a sampled section; The formula for calculating the NLI score is as follows: Where N is the number of sampled segments. The sequence number of the predicted text generated by the large language model. .

3. The text-assisted traffic prediction system according to claim 1, characterized in that: The specific working process of the time series feature extraction module includes the following tasks: (1) The time series preprocessing unit standardizes the historical network traffic, captures local semantics and long-range dependencies through the backbone model, and generates initial prediction results; (2) The multi-scale frequency decomposition unit decomposes the initial prediction result into features in different frequency domains through Discrete Fourier Transform (DFT). After being partitioned by a frequency selection function, it is converted into a multi-scale time series tensor in the time domain through Inverse Discrete Fourier Transform (IDFT). The calculation formula of the DFT is as follows: in Let be the initial prediction result in the time domain, and len be the total length of the network traffic vector of the initial prediction result. The time step takes a value from 0 to len−1. For frequency values, The imaginary unit, This represents the initial prediction result in the frequency domain. The frequency selection function is defined as: in For the frequency selection function of the i-th frequency band, Let be the lower limit frequency of the i-th frequency band. Let i be the upper limit frequency of the i-th frequency band. The value to be determined is the frequency value.

4. The text-assisted traffic prediction system according to claim 1 or 3, characterized in that: The specific working process of the multi-scale modality fusion module includes the following tasks: (1) Multi-scale convolutional units employ convolutional operations with differentiated dilation rates for temporal tensors of different frequency levels. High-frequency temporal tensors correspond to the rapid fluctuation characteristics of network traffic, and convolutional kernels with small dilation rates are used to accurately capture instantaneous change details. Low-frequency temporal tensors correspond to long-term stable trends, and convolutional kernels with large dilation rates are used to fully extract global trend information. The dilation rate of medium-frequency temporal tensors is between the two. Through differentiated convolution processing, the expressive power of temporal features in different frequency domains is enhanced, providing more targeted feature inputs for subsequent fusion. (2) The cross-attention fusion unit first performs linear projection transformation on the convolutional temporal tensors and text tensors at each scale. The temporal tensors at each scale are projected to generate query vectors Q, and the text tensors are projected to generate unified key vectors K and value vectors V. Then, attention weights are obtained by calculating the similarity between the query vectors and key vectors. These weights reflect the correlation strength between text semantic information and temporal features at each scale. Finally, the value vectors are weighted and aggregated based on the attention weights to obtain multi-scale fusion features that fuse text semantics and corresponding temporal features, thereby realizing the deep association between text and temporal data. The formula for calculating the convolution operation is: in Let be the weight value at the s-th position in the convolution kernel, where s is the index of the convolution kernel. Let be the value of the original time series tensor of the i-th frequency band at time step t. Let be the expansion rate of the i-th frequency band. This refers to the mapping time step after the convolution operation; The formula for calculating the fused features in the above steps is: , in, Let be the fused feature tensor obtained by fusing the temporal features and textual features of the i-th frequency band. For query vector dimensions, The query vector generated by linear projection of the time series tensor of the i-th frequency band. This is the transpose of the key vector generated by linear projection of the text tensor. It is a value vector generated by linear projection of the text tensor.

5. The text-assisted traffic prediction system according to claim 1 or 4, characterized in that: The specific working process of the adaptive prediction output module includes the following tasks: (1) The adaptive loss calculation unit calculates the initial prediction loss and the fusion prediction loss, and dynamically adjusts the loss value through the adaptive loss function. In the later stage of training, the fusion prediction loss is smoothly transitioned to the adaptive loss. (2) The projection output unit uses a 1×1 standard convolutional layer to map the multi-scale fused feature tensor to the target dimension, and outputs the predicted network traffic value within a set future time period. The formula for calculating the adaptive loss function is: in, The output loss value of the adaptive loss function. For the initial predicted loss, To fuse the prediction loss, Scale is the scaling factor for loss adjustment, used to adjust the magnitude of the loss function value, and Offset is the offset factor for loss adjustment, used to fine-tune the input baseline value of the loss function.

6. A text-assisted network traffic prediction method, characterized in that: The method includes the following steps: (1) Text processing and embedding generation: Historical network traffic data, corresponding domain background information and various external texts are input into the inference unit of the large language model. The predicted text conforming to a unified format is generated by the chain thinking CoT. The NLI score of each predicted text is calculated by the self-checking filter unit, and the phantom text with the score exceeding the threshold is removed. The filtered effective predicted text is input into the pre-trained BERT encoder, the model parameters are frozen for encoding processing, and a high-dimensional text tensor is generated. The zero vector is filled in the time step where there is no corresponding external text. (2) Temporal feature extraction and decomposition: The input historical network traffic is standardized to eliminate the difference in dimensions; the local semantics and long-range dependencies of the network traffic are captured by the selected backbone model to generate the initial prediction results; The initial prediction results are transformed to the frequency domain using Discrete Fourier Transform (DFT). Three non-overlapping sub-bands (high, medium, and low) are divided using a frequency selection function. Then, the signals of each sub-band are transformed back to the time domain using Inverse Discrete Fourier Transform (IDFT) to obtain a multi-scale time series tensor. (3) Multi-scale modal fusion: Perform convolution operations with different dilation rates on the temporal tensors of each scale to enhance the expression of features in different frequency domains; perform linear projection of the convolved temporal tensors and the text tensors to obtain query vectors, key vectors and value vectors; calculate attention weights and aggregate value vectors to achieve deep fusion of text semantics and temporal features of each scale, and obtain multi-scale fusion features. (4) Adaptive prediction output: Calculate the initial prediction loss and the fusion prediction loss respectively, and dynamically adjust the loss value through the adaptive loss function to achieve a smooth transition of loss during training; input the multi-scale fusion features into the standard convolutional layer for dimension mapping, and output the network traffic prediction value within the future set time period.

Images