Zero sample time sequence architecture design method based on multi-agent large language model
By employing a zero-sample time series architecture design method based on a multi-agent large language model, this approach addresses the issues of high computational cost and insufficient generalization ability in existing time series prediction models. It enables the automated design of efficient and low-cost time series prediction structures, making it suitable for resource-constrained real-world applications.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- EAST CHINA NORMAL UNIV
- Filing Date
- 2025-12-25
- Publication Date
- 2026-05-08
AI Technical Summary
Existing time series forecasting models suffer from high computational costs and insufficient generalization capabilities in terms of automated design and cross-task generalization, making them difficult to promote and apply, especially in resource-constrained scenarios.
We employ a multi-agent large language model, and achieve zero-shot time series architecture design through task-aware dataset profiling, similarity search space pruning, evolutionary enhancement modules, and external knowledge optimization. We combine operator agents and topological agents for search space pruning and introduce literature agents to enhance the generalization ability of the model.
It significantly reduces computational resource consumption, improves model design efficiency, enhances generalization ability on unseen datasets, and adapts to time series prediction tasks in complex scenarios.
Smart Images
Figure CN121998081A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of time series forecasting technology, and in particular to a zero-shot time series architecture design method based on a multi-agent large language model. Background Technology
[0002] With the widespread application of sensor and information acquisition technologies, numerous sensors have been deployed in social infrastructure such as transportation systems, power grids, and medical monitoring platforms to record various types of data that change over time. The resulting multidimensional time series data typically exhibits inherent correlation and is known as correlated time series. By analyzing and modeling historical correlated time series, accurate predictions of future values can be achieved, demonstrating significant application value in many practical scenarios such as traffic flow forecasting, power grid load scheduling, and patient health monitoring.
[0003] In recent years, deep learning methods have demonstrated significant advantages in related time series prediction tasks. The core of these models lies in the design of the spatio-temporal block (ST-block), which consists of spatial and temporal operators and can simultaneously capture the spatial correlation between sequences and the temporal dependence within a single sequence. However, existing Correlated Time Series (CTS) model architectures primarily rely on manual design, typically requiring researchers to possess extensive domain knowledge and design experience. This modeling process is time-consuming and costly, making it unsuitable for handling dynamically changing data environments and task requirements.
[0004] To overcome the limitations of manual design, automated modeling methods have emerged in recent years. Their core idea is to automatically discover high-performance ST-block structures within a predefined search space. This search space typically consists of spatial / temporal operators extracted from existing models and their typical topological connection patterns. Through search strategies such as reinforcement learning, evolutionary algorithms, or gradient optimization, the optimal network structure can be automatically searched within this space and embedded into the complete model for training and prediction. These methods achieve a certain degree of automated structure design and outperform traditional manual methods.
[0005] However, existing automated modeling methods still have two main problems:
[0006] (1) High computational and time costs. Existing methods generally require extensive model training and evaluation on the target dataset to explore the optimal structure in the search space. Due to the large search space and complex optimization objectives, the entire process consumes a lot of computational resources and time. Especially in CTS prediction tasks, model training itself is computationally intensive and time-consuming, making the search process more costly and difficult to promote in resource-constrained scenarios.
[0007] (2) Insufficient generalization ability. Meta-learning methods that have emerged in recent years attempt to achieve zero-shot structure search by learning transferable search strategies on multiple known datasets. However, these methods do not adequately model the structural features of new tasks or unseen datasets, resulting in poor generalization ability of the generated models on heterogeneous tasks. When the task distribution or data characteristics change, the prediction performance often drops significantly, limiting its applicability and stability in complex real-world scenarios.
[0008] In summary, existing time series forecasting models still face technical bottlenecks such as low efficiency, high resource consumption, and insufficient adaptability in terms of automated design and cross-task generalization. There is an urgent need for an efficient, low-cost automated modeling solution with good generalization capabilities. Summary of the Invention
[0009] This invention leverages the rich prior knowledge, powerful reasoning capabilities, and excellent scalability of Large Language Models (LLMs) to decompose complex ST-block design tasks through a multi-agent collaborative mechanism. Combining search space pruning and knowledge augmentation optimization modules, it achieves automated, intelligent, and efficient structural design. This method targets interconnected time-series data recorded by multiple sensors, utilizing the knowledge reasoning capabilities of LLMs to achieve zero-shot structural design of time-series prediction models without additional model training. Through a multi-agent collaborative mechanism, operator agents and topology agents are combined to perform search space pruning, improving model design efficiency. Simultaneously, literature agents and improvement agents are introduced to extract structural knowledge from academic literature and generate targeted improvement schemes, enhancing the model's generalization and adaptability to unknown tasks. This invention can achieve high-precision, high-efficiency automated design of time-series prediction structures while reducing computational resource consumption, demonstrating significant potential for widespread application.
[0010] The technical solution of this invention is as follows: A zero-shot time series architecture design method based on a multi-agent large language model, comprising the following steps:
[0011] Step 1: Use the task-aware dataset profiling module to find the most similar dataset from existing relevant time-series task benchmark datasets and characterize the target dataset;
[0012] Step 1.1: Calculate the similarity between the target dataset and multiple CTS benchmark datasets through language description and structural attribute analysis.
[0013] Step 1.2: Dataset similarity calculation and performance correlation analysis.
[0014] Step 2: The similarity search space pruning module prunes the complete search space based on the representation of the target dataset to obtain a high-performance subspace;
[0015] Step 2.1: Setting up operator agents and topology agents.
[0016] Step 2.2: Agent-guided search space pruning.
[0017] Step 3: In the high-performance subspace, use the evolutionary enhancement module to search and optimize the time series model architecture design; obtain the final determined time series model architecture.
[0018] Step 3.1: After completing the search space pruning, introduce a decision agent to construct the initial ST-block in the pruned space.
[0019] Step 3.2: External knowledge optimization; preprocess academic literature, parse and clean the documents, retain only the content related to the model architecture, segment the text into independent fragments and embed them into the vector space and store them in the knowledge base.
[0020] Step 3.3: ST-block integration and evaluation.
[0021] The task-aware dataset profiling module in step 1 is as follows:
[0022] The target dataset is represented using both linguistic descriptions and structural attributes.
[0023] The language description specifically refers to: using concise natural language text to describe the domain, feature semantics, and scale of the target time series dataset;
[0024] Specifically, the structural attributes are: automatically extracting statistical and signal features of the target time series dataset using the following feature engineering techniques:
[0025] 1) Smooth and reassemble the original target spatiotemporal sequence to remove local noise and obtain a denoised time series;
[0026] 2) On the denoised time series, a set of predefined operators are used to extract high-dimensional statistical feature vectors;
[0027] 3) On the high-dimensional statistical feature vector, the features most relevant to the prediction task are selected through significance testing;
[0028] 4) Sort by significance and select the top K most relevant features as the structural attribute representation of the target time series dataset.
[0029] A high-performance subspace is obtained by pruning the search space using the similarity search space pruning module from the complete search space;
[0030] The complete search space includes a subset of datasets similar to the target dataset from existing relevant time series task benchmark datasets, and the target dataset itself; each dataset has an effective model architecture, which is decomposed into operators and topology;
[0031] The operator agent is responsible for selecting a set of high-performance operators, and the topology agent is responsible for selecting high-quality topologies.
[0032] Operator agents and topological agents analyze the complete search space from three aspects: domain description, macro architecture, and operator and topology candidates. For each subset of the dataset, the existing model architecture is decomposed into operators and topologies, and the operator or topology types are clustered and ranked according to their average performance.
[0033] The top-ranked set of operators and topologies are extracted from the dataset subset as candidate operators and candidate topologies. The operator agent and the topology agent evaluate the candidate operators and candidate topologies based on task adaptability and generalization performance on the dataset subset. The operator agent and the topology agent exchange candidate results and cooperate to adaptively select. The operator agent refines the operators based on the feedback from the topology agent, and the topology agent refines the topologies based on the feedback from the operator agent. The final selection by the operator agent and the topology agent generates a pruned high-performance subspace.
[0034] The operator agent is a functional agent based on a large language model. Under zero training conditions, it selects a subset of operators that match the current relevant time series prediction task through a constrained reasoning process based on the statistical characteristics of the dataset and a predefined set of candidate operators, thereby effectively pruning the original search space.
[0035] The operator agent refines operators based on feedback from the topology agent. The topology agent is a functional agent based on a large language model. Given a candidate operator subspace, it generates or filters topological structures for relevant time series prediction tasks through a constrained reasoning process based on the structural characteristics of the dataset, thereby further narrowing the search space.
[0036] The evolutionary enhancement module designs and optimizes the time series model in the high-performance subspace, and its specific implementation is as follows:
[0037] 1) Initial architecture generation;
[0038] The decision-making agent generates an initial spatiotemporal block architecture within a high-performance subspace oriented towards the target task. Based on the definition of the complete search space, the description of the pruned high-performance subspace, the specific description of the target task, and design suggestions from operator agents and topology agents, the decision-making agent recommends candidate structures after comprehensive analysis. A vectorized literature knowledge base is established.
[0039] 2) Knowledge enhancement and optimization;
[0040] Establish an improvement agent and a document agent, which work together. The improvement agent analyzes the design of the initial spatiotemporal block architecture, identifies potential defects, and requests relevant knowledge from the document agent. The document agent retrieves and filters high-quality document fragments that are highly relevant to the task from the vectorized document knowledge base, and then feeds them back to the improvement agent. The improvement agent generates targeted optimization suggestions and adds a confidence score to each optimization suggestion.
[0041] 3) Iterative optimization and integration;
[0042] The decision agent receives optimization suggestions and their confidence scores from the improvement agent and incorporates them into the next round of spatiotemporal block architecture generation suggestions. The entire process is carried out in multiple iterations in a "generation-analysis-improvement" loop, and the spatiotemporal block architecture evolves continuously with each feedback. Finally, a time series model structure with both high performance and high interpretability is formed.
[0043] The vectorized literature knowledge base is established as follows: extract text content from time-series related research papers; automatically parse PDF files, clean up redundant information, and segment the text into fragments based on semantic continuity; then embed the fragments into a vector space to obtain the vectorized literature knowledge base.
[0044] The beneficial effects of this invention are as follows: This method has significant advantages in computational efficiency. No model training is required during the search phase; instead, ST-block design is completed through LLM inference and data-aware search space pruning, significantly reducing time and computational resource overhead, making it more suitable for resource-constrained real-world applications. This method exhibits stronger generalization ability on unseen datasets. Through a multi-agent collaborative LLM inference mechanism, combined with external structural knowledge from text modalities, cross-task transfer at the structural level can be achieved, alleviating the problem of insufficient generalization ability of existing zero-shot methods on heterogeneous CTS datasets. The multi-agent framework improves the inference quality of complex structural designs. By decomposing the ST-block design task into sub-tasks such as operator selection, topology modeling, and structural optimization, and assigning them to different agents, more stable and effective structural inference can be performed in large-scale discrete search spaces. Data-aware search space pruning improves the targeting of the search. Based on the characteristics of the target dataset, the selection of operators and topologies is guided, compressing the original search space into a high-performance subspace, reducing search complexity while improving the discovery efficiency of high-performance architectures. The introduction of a literature knowledge system enhances the robustness of the structural design. By retrieving and integrating high-quality structural information from CTS-related literature, effective architecture optimization was achieved without model training, further enhancing the adaptability of the method to diverse CTS tasks. Attached Figure Description
[0045] Figure 1 A framework diagram for designing a zero-shot time series architecture based on a multi-agent large language model;
[0046] Figure 2 Flowchart for pruning the search space;
[0047] Figure 3 Flowchart for the evolution enhancement module. Detailed Implementation
[0048] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.
[0049] like Figure 1 As shown, the method in this embodiment includes the following steps.
[0050] Step 1: Use the task-aware dataset profiling module to find the most similar dataset from existing relevant time-series task benchmark datasets and characterize the target dataset;
[0051] Step 1.1: Calculate the similarity between the target dataset and multiple CTS benchmark datasets through language description and structural attribute analysis. Since the goal of this invention is to automatically design high-performance ST-blocks on an unseen dataset, the first step is to perform a comprehensive feature analysis on the target dataset to extract key features, thereby supporting the subsequent structure search process. Structural patterns that exhibit strong generalization performance on one dataset often also achieve good results on other datasets with similar structural attributes. This finding provides a theoretical basis for dataset-level feature analysis in CTS tasks: by identifying similar datasets, effective reuse or adaptation of ST-block designs can be guided.
[0052] Therefore, this invention first designs a task-aware dataset profiling module to characterize the target dataset by identifying the most similar reference set in existing CTS benchmark datasets. This invention represents the dataset by combining the following two types of information:
[0053] (1) Language-based description. In order to incorporate semantic and contextual information into the dataset representation, this invention uses a concise text description for each dataset, denoted as P. desc These descriptions summarize the key attributes of the dataset, such as the data domain (e.g., transportation, energy), semantic features, and dataset size. Based on a comprehensive review of existing benchmark datasets and their documentation, these descriptions were manually compiled to ensure their accuracy and relevance. By capturing high-level information in natural language, these language-based descriptions provide a complementary modality for subsequent structure retrieval processes, thereby facilitating similarity calculation and benchmark matching.
[0054] (2) Structure-based attributes. In order to effectively capture the underlying statistical structure and time dependence contained in spatiotemporal data, this invention adopts a feature engineering process that combines automatic feature construction with salient feature screening.
[0055] The original spatiotemporal sequence of the target is reorganized and smoothed. Let the dataset be... ,in Indicates the number of time steps. Indicates the number of nodes. Represents the feature dimension. Each element... Represents a node At time step The feature vector is . To reduce local noise and enhance temporal stability, this invention uses a feature vector of length . A fixed sliding window is used, and the average value within each window is calculated:
[0056]
[0057] From each time series It automatically extracts a comprehensive set of statistical features to richly characterize time series behavior. Specifically, it uses a predefined set of operators. It contains various statistical and signal processing functions, thereby generating a high-dimensional feature vector:
[0058]
[0059] To identify the features most relevant to the prediction task, this invention evaluates each extracted feature. With target variable The statistical dependencies between them. For each feature, the following hypothesis tests are performed: Depending on the task type, tsfresh will automatically select the appropriate statistical test. For example, use ANOVA F-test or Kendall's rank correlation coefficient, and calculate the corresponding p-value:
[0060]
[0061] in, The observed test statistic value, This indicates the probability of observing this value under the null hypothesis. The smaller the value, the stronger the statistical association between the feature and the target variable. This invention retains the condition that... The characteristics, among which This is the significance threshold. The resulting set of relevant features is:
[0062]
[0063] To further reduce dimensionality and enhance generalization ability, this invention sorts the retained features according to the p-value and selects the top features with the strongest discriminative power. One characteristic:
[0064]
[0065] in Indicates the first The features corresponding to small p-values. This feature set is the final statistical representation of the dataset, denoted as . It summarizes the core statistical features extracted from time series data. To better capture the complex temporal dynamics and spatial structure of CTS data, this invention further introduces auxiliary information from both the time and spatial dimensions, denoted as... This enriches the dataset profile with spatiotemporal priors. Finally, all components are summed to form a complete feature representation of the dataset.
[0066]
[0067] Step 1.2: Dataset similarity calculation and performance correlation analysis. This invention will... As a signature for each dataset, pairwise similarity is calculated between datasets to identify similar datasets and characterize the target task. Specifically, for each pair of datasets... Distance between and performance differences during architecture migration Then, the Kendall rank correlation coefficient is used to measure the consistency of the two rankings:
[0068]
[0069] in, This indicates a sort based on the distance between the dataset images. This indicates a ranking based on architecture migration performance. A strong positive correlation (i.e.) This indicates that datasets with more similar profiles generally exhibit better transferability during structural transfer.
[0070] Step 2: The similarity search space pruning module prunes the complete search space based on the representation of the target dataset to obtain a high-performance subspace;
[0071] Step 2.1: Setting up Operator and Topology Agents. Given that topology and operators together constitute the basic components of an ST-block, this invention further decomposes the structural design task into two interdependent sub-problems: the selection of the operator set and the formulation of the topology connection scheme. This decomposition aligns with the hierarchical modeling principle commonly used in neural architecture search, reducing the complexity of the search space while maintaining design expressiveness and flexibility. By optimizing independently in these two key dimensions, the system can explore architectures in a more targeted and manageable subspace, thereby simultaneously improving the efficiency of the search process and the overall quality of the final architecture.
[0072] To address this, this invention designs an operator agent and a topology agent, corresponding to two key dimensions in the structural design. The operator agent is responsible for evaluating all operators in the search space and selecting a subset of promising candidate operators based on task-specific requirements. The topology agent focuses on evaluating various connection patterns and identifying topologies with strong representational capabilities and adaptability. It is important to note that although operators and topologies can be modeled as two relatively independent components, in practice they are inherently interdependent—optimal choices in isolation may lead to suboptimal performance when combined. Therefore, during the decision-making process, the two agents exchange evaluation information and collaboratively optimize their respective choices to ensure the consistency and effectiveness of the final architectural configuration.
[0073] Step 2.2: Agent-guided search space pruning. In the CTS domain, the design space of ST-blocks contains a large number of operator and topology combinations, making direct search inefficient and difficult to obtain high-quality structures. Therefore, this invention designs operator agents and topology agents to identify the most similar datasets based on task dataset features and historical CTS task information, and extracts highly transferable operator combinations and topology schemes from them.
[0074] The operator agent evaluates all operators and selects a candidate subset, while the topology agent evaluates connection patterns and selects the topology. The two agents exchange information and collaboratively optimize their choices during the decision-making process to ensure the consistency and overall performance of the final architecture in terms of both operators and topology.
[0075] Finally, the pruned set of operators and topology schemes are integrated to form a search space specific to the task, providing a foundation for subsequent model architecture design. For a detailed overview of this pruning process, please refer to [link to documentation / document]. Figure 2 .
[0076] Step 3: In the high-performance subspace, use the evolutionary enhancement module to search and optimize the time series model architecture design; obtain the final determined time series model architecture. For details of the overall process of this module, please refer to [link to module details]. Figure 3 .
[0077] Step 3.1: After pruning the search space, a decision agent is introduced to construct the initial ST-block within the pruned space. The architecture generated in the first round is usually not optimal, and fully training each candidate structure is costly. Therefore, an optimization agent and a literature agent are introduced to collaboratively optimize the design. The decision agent integrates information from multiple sources, including the complete search space definition, the description of the pruned subspace, the target task specification, and the candidate operator set and topology set provided by the operator agent and topology agent. By querying similar CTS benchmark datasets, existing ST-block designs that satisfy the constraints are retrieved, providing practical references and validating structural patterns for the initial architecture recommendation.
[0078] Step 3.2: External Knowledge Optimization; Preprocessing of academic literature involves parsing and cleaning the documents, retaining only content relevant to the model architecture, and segmenting the text into independent fragments, which are then embedded into a vector space and stored in the knowledge base. First, the external knowledge extracted from academic papers is preprocessed. Specifically, after extracting the original text from the PDF file, document parsing and text cleaning are performed to remove redundant elements such as titles and figure captions, retaining only the core content related to the model architecture design. Next, the text is segmented into coherent fragments based on length and semantic coherence. Each fragment is then embedded into a vector space and stored as an independent document in the knowledge base to facilitate downstream retrieval and enhance the generation process.
[0079] This invention introduces a collaborative optimization agent and a literature agent. To support the optimization process, an optimization agent and a literature agent are introduced to collaboratively extract relevant knowledge from a vector-based database and provide constructive suggestions to the decision agent. The optimization agent first receives the design task assigned to the decision agent and the resulting initial architectural proposal. It then conducts a comprehensive review and analysis of the proposed structure. After identifying potential flaws or limitations in the current design, the optimization agent proactively requests external knowledge from the literature agent to guide and inspire structural improvements. The literature agent then retrieves relevant information from a vectorized knowledge base constructed from academic literature. It performs a quality assessment of the retrieved content, focusing on architectural insights, task relevance, and practical utility. Based on this assessment, the literature agent filters out noisy or overly generalized descriptions, retaining only high-quality knowledge fragments that provide specific and context-appropriate design inspiration.
[0080] The optimization agent combines selected knowledge fragments to generate targeted modification suggestions, along with a confidence score. The agent integrates these selected knowledge fragments with the current architecture, comprehensively analyzing potential directions for structural enhancement. Importantly, each suggestion is assigned a confidence score. This confidence score is derived by comprehensively considering the consistency of supporting literature, the credibility of the sources, and the similarity between the current task and reference cases. This mechanism helps reduce the risk of illusions during the optimization process and improves the reliability and applicability of the generated suggestions.
[0081] Optimization suggestions and their associated confidence scores are fed back to the decision-making agent. Through multiple rounds of "generate-analyze-optimize" iterations, the initial ST-block architecture is gradually improved by incorporating high-quality external knowledge and optimization feedback, ultimately resulting in a high-performance ST-block design.
[0082] Step 3.3: ST-block integration and evaluation.
[0083] After obtaining the final ST-block, it is integrated into the neural network according to its operator design and topological connection structure, followed by a full training cycle to evaluate the model's performance metrics. We evaluate the performance of our proposed method versus existing methods on multiple metrics such as MAE, RMSE, and MAPE on five public datasets (METR-LA, PEMS03, PEMS04, PEMS07, PEMS08).
[0084] Table 1 presents our evaluation of the proposed method versus existing methods (AGCRN, PDFormer, LightST, AutoCTS++, FACTS) on five public datasets (METR-LA, PEMS03, PEMS04, PEMS07, PEMS08) across multiple metrics, including MAE, RMSE, and MAPE, covering different prediction settings. Experimental results include short-term (P-12 / Q-12) and long-term (P-48 / Q-48) prediction ranges to comprehensively evaluate the generalization ability of our method. We use bold to indicate the best performance and underline the second-best results for each setting. Our method consistently outperforms existing manual and automated methods on almost all evaluation metrics.
[0085] Table 1 Comparison of metrics for different methods across different datasets
[0086]
Claims
1. A zero-shot time series architecture design method based on a multi-agent large language model, characterized in that, Includes the following steps: Step 1: Use the task-aware dataset profiling module to find the most similar dataset from existing relevant time-series task benchmark datasets and characterize the target dataset; Step 2: The similarity search space pruning module prunes the complete search space based on the representation of the target dataset to obtain a high-performance subspace; Step 3: In the high-performance subspace, use the evolutionary enhancement module to search and optimize the time series model architecture design; obtain the final determined time series model architecture.
2. The zero-shot time series architecture design method based on a multi-agent large language model according to claim 1, characterized in that, The task-aware dataset profiling module in step 1 is as follows: The target dataset is represented using both linguistic descriptions and structural attributes. The language description specifically refers to: using concise natural language text to describe the domain, feature semantics, and scale of the target time series dataset; Specifically, the structural attributes are: automatically extracting statistical and signal features of the target time series dataset using the following feature engineering techniques: 1) Smooth and reassemble the original target spatiotemporal sequence to remove local noise and obtain a denoised time series; 2) On the denoised time series, a set of predefined operators are used to extract high-dimensional statistical feature vectors; 3) On the high-dimensional statistical feature vector, the features most relevant to the prediction task are selected through significance testing; 4) Sort by significance and select the top K most relevant features as the structural attribute representation of the target time series dataset.
3. The zero-shot time series architecture design method based on a multi-agent large language model according to claim 1, characterized in that, A high-performance subspace is obtained by pruning the search space using the similarity search space pruning module from the complete search space; The complete search space includes a subset of datasets similar to the target dataset from existing relevant time series task benchmark datasets, and the target dataset itself; each dataset has an effective model architecture, which is decomposed into operators and topology; The operator agent is responsible for selecting a set of high-performance operators, and the topology agent is responsible for selecting high-quality topologies. Operator agents and topological agents analyze the complete search space from three aspects: domain description, macro architecture, and operator and topology candidates. For each subset of the dataset, the existing model architecture is decomposed into operators and topologies, and the operator or topology types are clustered and ranked according to their average performance. The top-ranked set of operators and topologies are extracted from the dataset subset as candidate operators and candidate topologies. The operator agent and the topology agent evaluate the candidate operators and candidate topologies based on task adaptability and generalization performance on the dataset subset. The operator agent and the topology agent exchange candidate results and cooperate to adaptively select. The operator agent refines the operators based on the feedback from the topology agent, and the topology agent refines the topologies based on the feedback from the operator agent. The final selection by the operator agent and the topology agent generates a pruned high-performance subspace.
4. The zero-shot time series architecture design method based on a multi-agent large language model according to claim 3, characterized in that, The operator agent is a functional agent based on a large language model. Under zero training conditions, it selects a subset of operators that match the current relevant time series prediction task through a constrained reasoning process based on the statistical characteristics of the dataset and a predefined set of candidate operators, thereby effectively pruning the original search space.
5. The zero-shot time series architecture design method based on a multi-agent large language model according to claim 3, characterized in that, The operator agent refines operators based on feedback from the topology agent. The topology agent is a functional agent based on a large language model. Given a candidate operator subspace, it generates or filters topological structures for relevant time series prediction tasks through a constrained reasoning process based on the structural characteristics of the dataset, thereby further narrowing the search space.
6. The zero-shot time series architecture design method based on a multi-agent large language model according to claim 1, characterized in that, The evolutionary enhancement module designs and optimizes the time series model in the high-performance subspace, and its specific implementation is as follows: 1) Initial architecture generation; The decision-making agent generates an initial spatiotemporal block architecture within a high-performance subspace oriented towards the target task. Based on the definition of the complete search space, the description of the pruned high-performance subspace, the specific description of the target task, and design suggestions from operator agents and topology agents, the decision-making agent recommends candidate structures after comprehensive analysis. A vectorized literature knowledge base is established. 2) Knowledge enhancement and optimization; Establish and improve intelligent agents and document intelligent agents, which work together; The improved agent analyzes the initial spatiotemporal block architecture design, identifies potential defects, and requests relevant knowledge from the literature agent. The literature agent retrieves and filters high-quality, task-relevant literature fragments from the vectorized literature knowledge base, and then feeds them back to the improved agent. The improved agent generates targeted optimization suggestions and adds a confidence score to each optimization suggestion. 3) Iterative optimization and integration; The decision agent receives optimization suggestions and their confidence scores from the improvement agent and incorporates them into the next round of spatiotemporal block architecture generation suggestions. The entire process is carried out in multiple iterations in a "generation-analysis-improvement" loop, and the spatiotemporal block architecture evolves continuously with each feedback. Finally, a time series model structure with both high performance and high interpretability is formed.
7. The zero-shot time series architecture design method based on a multi-agent large language model according to claim 6, characterized in that, The vectorized literature knowledge base is established as follows: extract text content from time-series related research papers; automatically parse PDF files, clean up redundant information, and segment the text into fragments based on semantic continuity; then embed the fragments into a vector space to obtain the vectorized literature knowledge base.