LLM-enabled multi-agent intermittent demand forecasting method and system

Abstract

The application discloses an LLM-enabled multi-agent intermittent demand prediction method and system, which comprises a data monitoring stage capable of converting an input original intermittent demand sequence into a high-quality data set; a feature extraction stage for converting the high-quality data set into a feature combination by analyzing data characteristics through recursive reasoning, autonomously planning steps and strategies of feature engineering, and coordinating feature calculation and fusion tools; a pattern classification stage for converting the feature combination into a demand pattern category according to the feature structure matching a clustering algorithm, and then generating readable labels based on the clustering algorithm result statistics characteristics; and a model selection stage for finally outputting a model recommendation scheme matched for each demand pattern by searching a structured model knowledge base through enhanced generation technology, combining data features and demand pattern categories for multi-dimensional deduction. The application can balance the explainability in the intermittent demand prediction scene, and the decision-making process is transparent and the prediction accuracy is high.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent manufacturing, and in particular to an LLM-enabled method and system for predicting intermittent demand among multiple agents. Background Technology

[0002] In the field of intelligent manufacturing, the long service life of large equipment makes spare parts management a prominent challenge for automotive and aerospace companies. Spare parts supply assurance is a crucial link in intelligent manufacturing, directly impacting equipment operating efficiency and downtime losses. Compared to finished product demand, spare parts demand exhibits a distinct long-tail distribution and irregular fluctuations: a wide variety of types with few samples of any single type, often showing periods of zero demand interspersed with sporadic bursts. Demand differences between different factories, equipment, and processes are also significant. This makes demand forecasting not only crucial for inventory optimization and stockout risk control, but also profoundly impacts service levels, operational response efficiency, and maintenance planning, becoming a core link between the operational efficiency and financial performance of intelligent manufacturing.

[0003] While businesses often stockpile high-priced inventory to prevent downtime, the extreme sparsity and irregularity of demand still lead to repeated instances of "being unprepared when needed, or having inventory but not using it." Traditional forecasting methods often struggle to balance scalability and interpretability, while directly applying Large Language Models (LLMs) cannot effectively analyze the inherent structure of such sparse data.

[0004] Currently, the industry urgently needs a forecasting method and system that can balance interpretability, transparent decision-making process, and high forecasting accuracy in intermittent demand forecasting scenarios. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides an LLM-enabled multi-agent intermittent demand forecasting method and system. This method can significantly outperform classical statistical methods and machine learning models in terms of prediction accuracy while maintaining the transparency of the decision-making process, and also has a lower error rate. It has good overall advantages in intermittent demand forecasting scenarios.

[0006] In some embodiments of this application, a multi-agent intermittent demand forecasting method and system based on LLM is constructed. This system abandons the traditional hierarchical architecture and adopts a system flexibly scheduled by a state diagram. It coordinates four specialized agents—data monitoring, feature extraction, pattern classification, and model selection—to form a dynamic collaborative process. The data monitoring agent, leveraging semantic reasoning capabilities, accurately identifies real anomalies in business fluctuations, ensuring the quality of input data. The feature extraction agent, through context-aware input engineering, transforms raw demands into adaptive representations that integrate statistical and deep features. The pattern classification agent analyzes clustering results, generating semantic labels easily understood by the business, categorizing complex demand patterns into clear semantic categories. The model selection agent, combining a retrieval-enhanced generation (RAG) mechanism with an adaptive feedback loop, invokes an external expert knowledge base to match the optimal prediction model for each type of demand. Through the deep integration of semantic understanding capabilities and statistical methods, we have constructed a multi-agent intermittent demand forecasting system with autonomous feature extraction, interpretable pattern recognition, and knowledge-guided model selection capabilities, ultimately establishing a new balance between prediction performance and decision transparency.

[0007] To achieve the above objectives, some embodiments of this application provide an LLM-enabled multi-agent intermittent demand prediction method, which includes the following stages: data monitoring, feature extraction, pattern classification, and model selection; During the data monitoring phase, the data monitoring agent relies on the semantic parsing and context discrimination capabilities of the large language model to intelligently schedule data cleaning tools, transforming the input raw intermittent demand sequence containing mixed noise into a high-quality dataset. In the feature extraction stage, the feature extraction agent uses a large language model as its strategy brain, dynamically analyzes the data characteristics of high-quality datasets through iterative reasoning, autonomously plans the steps and strategies of feature engineering, and coordinates feature computing and fusion tools to transform high-quality datasets into feature combinations that can capture complex patterns. In the pattern classification stage, the pattern classification agent relies on the strategy selection and semantic generation function of the large language model. The pattern classification agent first matches a suitable clustering algorithm according to the feature structure of the feature combination, and then generates readable labels based on the statistical characteristics of the clustering algorithm results, transforming the feature combination into a required pattern category with business semantics. In the model selection phase, the model selection agent uses a large language model as the reasoning engine that integrates external knowledge. It calls the structured model knowledge base through retrieval enhancement generation technology, combines data features and demand pattern categories to perform multi-dimensional inference, and finally outputs a model recommendation scheme that matches each type of demand pattern.

[0008] In some embodiments, prior to the data monitoring phase, the prediction method further includes establishing a modular shared toolkit, which is invoked during the data monitoring phase, feature extraction phase, pattern classification phase, and model selection phase, and includes the tools required for the data monitoring phase, feature extraction phase, pattern classification phase, and model selection phase.

[0009] In some embodiments, the modular shared toolkit includes an anomaly monitoring and handling module, a data preprocessing module, a data standardization module, a statistical feature calculation module, a time series feature calculation module, a TSfresh time series feature calculation module, SBC classification, an autoencoder feature extraction module, recursive feature elimination, SHAP feature importance analysis, K-means clustering, hierarchical clustering, cluster evaluation, historical storage, LightGBM model prediction, TBS model prediction, RAG TOP-K model selection, and operation report generation.

[0010] In some embodiments, prior to the data monitoring step, the prediction method further includes establishing a standardized invocation protocol. The standardized invocation protocol is used to construct a dynamic closed-loop mechanism for perception, decision-making, and execution. The data monitoring agent, feature extraction agent, pattern classification agent, and model selection agent, along with the modular shared tool library, follow the standardized invocation protocol. Furthermore, the data monitoring agent, feature extraction agent, pattern classification agent, and model selection agent form a whole through the standardized invocation protocol.

[0011] In some embodiments, the scheduling of the collaborative process of multiple agents within the prediction method is handled by a state graph execution engine built on the LangGraph framework. The engine positions the large language model as the intelligent decision-making hub, where the multiple agents include a data monitoring agent, a feature extraction agent, a pattern classification agent, and a model selection agent.

[0012] In some embodiments, the collaborative process of multiple agents within the prediction method is a directed state graph. The engine has a global short-term memory. During the data monitoring stage, feature extraction stage, pattern classification stage, and model selection stage, multiple agents write processing data and running status to the global short-term memory in real time to achieve information sharing and backtracking.

[0013] In some embodiments, when the modular shared tool library is invoked by the data monitoring agent, feature extraction agent, pattern classification agent, and model selection agent, the large language model selects a specific tool through logical reasoning, generates standardized instructions to drive the execution of the modular shared tool library, and updates the state based on the execution results to enter the next round of decision-making.

[0014] In some embodiments, the prediction method includes an adaptive feedback integration mechanism that supports the dynamic path adjustment of the multi-agent collaborative process of the prediction method. When the prediction error exceeds the limit in the model selection step, the engine rolls back the collaborative process status to the feature extraction node through the conditional transition path. If the received signal indicates that there are insufficient lagging features, the large language model will re-plan, and the adaptive feedback integration mechanism will automatically generate dynamic instructions to guide feature extraction by adding short-term lagging features.

[0015] Some embodiments of this application also provide an LLM-enabled multi-agent intermittent demand prediction system, including a data monitoring agent, a feature extraction agent, a pattern classification agent, and a model selection agent. The data monitoring agent is used to perform the semantic parsing and context discrimination capabilities based on a large language model as described in any of the above embodiments, intelligently scheduling data cleaning tools to transform the input raw intermittent demand sequence containing mixed noise into a high-quality dataset. The feature extraction agent is used to perform the data characteristics of the high-quality dataset dynamically analyzed through iterative reasoning using a large language model as the strategy brain, autonomously planning the steps and strategies of feature engineering, and flexibly coordinating feature calculation and fusion tools as described in any of the above embodiments, to transform the high-quality demand sequence into a high-quality dataset. The process involves transforming a high-quality dataset into feature combinations capable of capturing complex patterns; a pattern classification agent performs strategy selection and semantic generation based on a large language model, as described in any of the above embodiments. It first matches a suitable clustering algorithm based on the feature structure of the feature combinations, then generates readable labels based on the statistical characteristics of the clustering algorithm results, transforming the feature combinations into demand pattern categories with business semantics; a model selection agent performs the process described in any of the above embodiments, using a large language model as an inference engine that integrates external knowledge. It calls a structured model knowledge base through retrieval enhancement generation technology, combines data features with the demand pattern categories for multi-dimensional deduction, and finally outputs a model recommendation scheme matching each demand pattern category.

[0016] Some embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the prediction method described above.

[0017] This application, in some embodiments, specifically relates to an LLM-enabled multi-agent intermittent demand forecasting method and system, fundamentally changing the application paradigm of large language models: instead of using them as direct prediction tools, they serve as the intelligent scheduling hub of the prediction system, constructing a multi-agent collaborative system from general scheduling to specific execution: designing four types of agents based on the ReAct architecture, autonomously coordinating through standardized collaboration protocols to achieve a closed-loop operation from data input to prediction output; developing a context-aware semantic data processing mechanism: proposing an innovative input engineering scheme to transform raw time-series data into structured statistical meta-features, enabling the large language model to identify abnormal fluctuations based on business semantics and dynamically adjust feature processing strategies; creating an interpretable pattern recognition path based on semantic generation: establishing a semantic interpretation process for clustering results, generating a label system that conforms to business cognition through the large language model, achieving deep integration of numerical features and domain knowledge; forming a knowledge-enhanced adaptive decision-making mode: constructing a model selection mechanism that integrates retrieval-enhanced generation (RAG) and multi-agent feedback, enabling the prediction system to both call external expert knowledge bases and continuously optimize itself based on actual prediction results. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the multi-agent collaboration process of an LLM-enabled multi-agent intermittent demand prediction system provided in some embodiments of this application. Figure 2 for Figure 1 The diagram shows the context-aware adaptive input engineering process of the prediction system for large language models. Figure 3 for Figure 1 The diagram shows a code listing for an adaptive feedback integration mechanism in an LLM-enabled multi-agent intermittent demand forecasting system. Figure 4 for Figure 1 The diagram shows the workflow of a data monitoring agent in an LLM-enabled multi-agent intermittent demand forecasting system. Figure 5 for Figure 1 The diagram shows the workflow of a feature extraction agent in an LLM-enabled multi-agent intermittent demand forecasting system. Figure 6 for Figure 5 The diagram shows a code listing illustrating the working process of the feature extraction agent. Figure 7 for Figure 5 The diagram shows a code listing illustrating the adaptive feature selection and fusion process within the intelligent feature extraction system. Figure 8 for Figure 5The diagram shows the steps of the feature extraction intelligent intracellular adaptive feature selection and fusion process; Figure 9 for Figure 5 The diagram shows the steps involved in the feature extraction agent's workflow. Figure 10 for Figure 1 The diagram shows the workflow of a pattern classification agent in an LLM-enabled multi-agent intermittent demand forecasting system. Figure 11 for Figure 10 The diagram shows a code listing for a pattern classification agent. Figure 12 for Figure 10 The diagram shown illustrates the working process of the large language model within the pattern classification intelligent system. Figure 13 for Figure 1 The diagram shows the workflow of model agent selection in an LLM-enabled multi-agent intermittent demand forecasting system. Figure 14 A schematic diagram of the execution code for a JSON-based structured interaction protocol provided in an embodiment of this application; Figure 15 This is a schematic diagram of an electronic device in some embodiments of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0021] In the field of intelligent manufacturing, the long service life of large equipment makes spare parts management a prominent challenge for automotive and aerospace companies. Ensuring spare parts supply is a crucial aspect of intelligent manufacturing, directly impacting equipment operating efficiency and downtime losses. While companies often stockpile high-priced inventory to prevent downtime, the extreme sparsity and irregularity of demand still lead to repeated instances of "neither needing to prepare nor preparing but not using." Compared to finished product demand, spare parts demand exhibits a distinct long-tail distribution and irregular fluctuations: SKUs (Stock Keeping Units, unique identifiers assigned by merchants for refined management and tracking) are numerous, but individual categories have few samples, often resulting in long periods of zero demand interspersed with sporadic bursts. Demand differences between different factories, equipment, and processes are also significant. This makes demand forecasting not only crucial for inventory optimization and stockout risk control but also profoundly impacts service levels, operational response efficiency, and maintenance planning, becoming a core link between the operational efficiency and financial performance of intelligent manufacturing.

[0022] Traditional prediction methods often struggle to balance scalability and interpretability, while directly applying Large Language Models (LLMs) cannot effectively analyze the intrinsic structure of such sparse data.

[0023] like Figure 1 and Figure 2 As shown, this invention provides an LLM-enabled multi-agent intermittent demand forecasting system, which can be used to predict sparse, intermittent spare parts demand. It includes a data monitoring agent, a feature extraction agent, a pattern classification agent, and a model selection agent. Each agent makes autonomous decisions by comprehensively analyzing upstream output, preset instructions, and historical experience, flexibly calling upon tool modules, and continuously looping a "reasoning-execution-feedback" closed-loop process, forming an end-to-end adaptive intelligent forecasting system (a collaborative process of multiple agents) from data monitoring to model optimization. Here, an agent represents an entity within the corresponding system that possesses autonomy, adaptability, and interactivity, capable of perceiving changes in the environment (e.g., through sensors or data input), making judgments and decisions based on its learned knowledge and algorithms, and then executing actions to influence the environment or achieve predetermined goals.

[0024] Specifically, in the embodiments of this application, the data monitoring agent can act as the "quality control center" of the prediction system, responsible for processing the original intermittent demand sequence containing mixed noise. Transform into a clean and well-organized high-quality dataset. Relying on the semantic parsing and context discrimination capabilities of large language models, it can accurately identify anomalies and normal business fluctuations, thereby intelligently scheduling data cleaning tools to ensure the quality of input data.

[0025] The feature extraction agent can act as the "feature engineer" of the prediction system, responsible for normalizing the high-quality data set. Transformed into feature combinations that can keenly capture complex patterns The feature extraction agent uses a large language model as its strategic brain, dynamically analyzes data characteristics through iterative reasoning, autonomously plans the steps and strategies of feature engineering, and flexibly coordinates feature calculation and fusion tools to transform the original requirements into an adaptive representation that integrates statistical features and deep features.

[0026] Pattern classification agents can act as "pattern decoders" within a prediction system, responsible for combining numerical features. Transform into a category of requirement patterns with business semantics Among them, the large language model plays a dual role in clustering strategy selection and semantic generation: first, it matches a suitable clustering algorithm based on the feature structure, and then generates readable labels such as "high-frequency stable consumption type" based on the statistical characteristics of the clustering results, so as to classify complex demand patterns into clear semantic categories and realize the connection between data patterns and business cognition.

[0027] The model selects an agent as the "decision terminal" of the prediction system, responsible for classifying each type of demand pattern. Matching the optimal prediction model Among them, the large language model serves as the reasoning engine that integrates external knowledge. It calls upon the structured model knowledge base through retrieval-enhanced generation technology, combines data features and demand pattern categories to perform multi-dimensional inference, and finally outputs a model recommendation solution with explanations and confidence scores.

[0028] It is understood that, in this embodiment, the general capabilities of the large language model are concretized into four ReAct agents with clearly defined roles and dynamic collaborative operation. These agents make autonomous decisions by comprehensively analyzing upstream outputs, preset instructions, and historical experience, flexibly calling upon specialized modules in a shared tool library, and continuously looping the "reasoning-execution-feedback" closed-loop process. These four types of agents form an end-to-end LLM-enabled multi-agent intermittent demand prediction system, from the initial intermittent demand sequence to the final decision. In this framework, the large language model does not directly perform numerical prediction but undertakes the key functions of core reasoning and intelligent scheduling. Relying on its semantic parsing, logical deduction, and knowledge construction capabilities, the large language model drives the iterative optimization of each agent until the task is accomplished through continuous state analysis, action planning, and effect evaluation. The internal operating mechanisms, tool call links, and collaborative feedback systems of each agent together constitute the dynamic collaborative core of the framework. The large language model provides the entire prediction system with the core driving force for intelligent evolution and autonomous adaptation. By deeply integrating semantic understanding capabilities with statistical methods, a multi-agent collaborative system with autonomous feature extraction, interpretable pattern recognition, and knowledge-guided model selection capabilities was constructed, ultimately establishing a new balance between prediction performance and decision transparency.

[0029] In some embodiments, the LLM-enabled multi-agent intermittent demand forecasting system method provided in this application may include a data monitoring stage, a feature extraction stage, a pattern classification stage, and a model selection stage. The data monitoring agent, relying on the semantic parsing and contextual discrimination capabilities of a large language model, intelligently schedules data cleaning tools to transform the input raw intermittent demand sequence containing mixed noise into a high-quality dataset; this process is denoted as the data monitoring stage. The feature extraction agent, using the large language model as its strategy engine, dynamically analyzes the data characteristics of the high-quality dataset through iterative reasoning, autonomously plans the steps and strategies of feature engineering, and flexibly coordinates feature computation and fusion tools to transform the high-quality dataset... The process of transforming feature combinations into features capable of capturing complex patterns is denoted as the feature extraction stage. The pattern classification agent, relying on the strategy selection and semantic generation functions of the large language model, first matches a suitable clustering algorithm based on the feature structure of the feature combinations, and then generates readable labels based on the statistical characteristics of the clustering algorithm results. The process of transforming feature combinations into demand pattern categories with business semantics is denoted as the pattern classification stage. The model selection agent uses the large language model as an inference engine that integrates external knowledge. Through retrieval enhancement generation technology, it calls the structured model knowledge base, combines data features with the demand pattern categories to perform multi-dimensional inference, and finally outputs a model recommendation scheme matching each demand pattern category. This process is denoted as the model selection stage.

[0030] For example, a modular shared tool library can be established before the data monitoring step. This tool library is invoked during the data monitoring, feature extraction, pattern classification, and model selection stages, covering the tools required for each stage. The tool library integrates full-chain functional modules covering data monitoring, feature extraction, pattern classification, model selection, and performance evaluation. Each tool is encapsulated as a callable unit with a unified interface and independent functionality. Clearly defined functional semantics and calling specifications enable the large language model to accurately identify and schedule these tools. Through this design, the originally complex task of spare parts demand prediction is deconstructed into a standardized sequence of steps that the large language model can plan and the tool library can execute, thus establishing an efficient collaborative mechanism between the general semantic understanding of the large language model and the specialized computational capabilities of the tool library.

[0031] As shown in Table 1 below, in some examples, the modular shared toolkit may include modules for anomaly monitoring and handling, data preprocessing, data standardization, calculating statistical features, calculating time series features, TSfresh calculating time series features, SBC classification, autoencoder feature extraction, recursive feature elimination, SHAP feature importance analysis, K-means clustering, hierarchical clustering, cluster evaluation, historical storage, LightGBM model prediction, TBS model prediction, RAG TOP-K model selection, and generating operation reports, among other tools.

[0032] Table 1 Overview of Core Multi-Agent Tools

[0033] In some embodiments, a standardized calling protocol can be built before the data monitoring step. The standardized calling protocol is used to build a dynamic closed-loop mechanism for perception, decision-making and execution. The calling of the data monitoring agent, feature extraction agent, pattern classification agent and model selection agent and the modular shared tool library follows the standardized calling protocol, and the data monitoring agent, feature extraction agent, pattern classification agent and model selection agent form a whole through the standardized calling protocol.

[0034] Understandably, in this embodiment, a multi-agent collaborative system is constructed, encompassing general scheduling and specialized execution. All agents autonomously collaborate through a standardized invocation protocol, achieving a closed-loop operation from data input to predicted output. Furthermore, the interaction between the multiple agents and the tool library follows a unified standardized invocation protocol, with the large language model playing a core decision-making role. In each iteration, the large language model parses the current task state, selects the tool to be invoked (a specific function) through logical reasoning, generates standardized instructions to drive the tool library's execution, and updates its cognitive state based on the execution results to proceed to the next round of decision-making. This protocol ensures the standardized uniformity of the agents' behavioral patterns and provides reliable support for system-level coordination of complex tasks. In other words, the modular shared tool library and standardized invocation protocol in this application form the technological foundation supporting end-to-end collaboration.

[0035] In summary, by establishing a standardized invocation protocol with a large language model as the decision-making center, we have built a highly efficient and collaborative underlying interaction framework for the four intelligent agents: data monitoring, feature extraction, pattern classification, and model selection. Within this framework, intelligent agents with distinct specializations are organically integrated to jointly construct a complete end-to-end intelligent prediction system driven by a large language model.

[0036] like Figure 1 and Figure 2As shown, in some embodiments, the scheduling of the collaborative process of the LLM-enabled multi-agent intermittent demand prediction method is handled by a state graph execution engine built on the LangGraph framework, which positions the large language model as the intelligent decision-making center. Unlike the traditional linear thought chain execution mode, this framework models the collaborative process of the prediction method as a directed state graph: each agent acts as a state node, and the four core agents form an end-to-end information processing link according to the prediction task logic. Their interaction relationships constitute the state transition path. The large language model dynamically plans the execution order of each node based on the real-time context, improving the flexibility of multi-agent collaboration. Adhering to the "divide and conquer" principle, this framework uses the state graph engine for overall scheduling, constructing a functionally modular and logically integrated system architecture.

[0037] In some embodiments, the engine may also have a global short-term memory (memory module) into which all agents write processing data and running status in real time, enabling system-level information sharing and backtracking. This design not only retains the output results of each stage but also records execution metadata, including error logs and retries. Crucially, this mechanism supports dynamic path adjustment: for example, when the model selection node detects that the prediction error exceeds the limit, the engine can roll back the state to the feature extraction node via a conditional transition path, rather than directly terminating the process, greatly improving system fault tolerance and collaborative efficiency.

[0038] In some embodiments, the multi-agent intermittent demand forecasting system includes an adaptive feedback integration mechanism to form a closed-loop self-optimization capability. This mechanism supports dynamic path adjustment in the collaborative process of the forecasting method. Under this mechanism, feedback signals are no longer simple error reports, but structured signals containing specific problem diagnoses and optimization suggestions. These signals are recorded in the global state. In this field, when the upstream agent is reactivated through the state graph loop mechanism, this feedback signal will be read first to adjust the execution strategy.

[0039] For example, such as Figure 1 and Figure 3 As shown, when the model-selected agent detects a significant prediction bias, the system automatically generates dynamic instructions to guide the feature extraction agent to add short-term lag features. This mechanism enables the system to evolve from a fixed process into a dynamic agent that can continuously self-optimize based on downstream performance.

[0040] For example, please refer to again Figure 1 and Figure 2 The collaborative process of an LLM-enabled multi-agent intermittent demand forecasting system can be summarized as follows: Please refer to the following: Figure 1 and Figure 4S1: In the data monitoring phase, as the perception front end of the LLM-enabled multi-agent collaborative intermittent demand forecasting system, relying on the semantic parsing and context discrimination capabilities of the large language model, it executes a closed-loop data cleaning process with semantic awareness. This process cleans the raw, unstructured, noisy, and intermittent demand sequences. Transform into a highly reliable, standardized, and cleaned high-quality dataset. It combines statistical anomaly detection with the semantic reasoning capabilities of a large language model to distinguish between genuine anomalies and normal business fluctuations. Its operational logic adopts a cyclical structure of "perception-reasoning-execution-observation".

[0041] like Figure 4 As shown, Step 1: For example, the data monitoring phase may include a quality scanning and evaluation process. The data monitoring agent can first analyze the input raw data (raw intermittent demand sequence). Perform a rapid quality scan to calculate the proportion of missing values ​​and the proportion of zero values ​​in the original intermittent demand series data, and obtain a basic quality score. :

[0042] in This represents the number of missing values. This represents the number of duplicate records. This represents the total length of the sequence.

[0043] For example, the preset threshold is If the basic quality score Then, the Step 2 deep anomaly monitoring process will be executed; Step 2: The data monitoring phase can include a deep anomaly detection process. The data monitoring agent can call the detect_anomalies_statistical() tool to combine statistical algorithms with the semantic reasoning capabilities of large language models for joint diagnosis. Specifically, firstly, an improved Isolation Forest algorithm is used for each data point. Calculate anomaly score Its definition is as follows:

[0044] in, Indicates sample Expected path length in all isolated trees Its actual path length, This is in the case of a sample size of The normalization constant for the path length.

[0045] For all satisfying Suspected anomalies Among them, the abnormal scoring threshold It can be set to 0.65. This will launch the large language model semantic reasoning engine for confirmation; the input for semantic reasoning is structured context. ,include: front and back The demand value at a given point in time, and the SKU (stock keeping unit, a unique identifier assigned by the merchant to a product for refined management and tracking) of that spare part in the past Relevant business events within the month, such as maintenance work orders, production plan changes, and promotional calendar events. The large language model analyzes this context and outputs confirmation markers. and the corresponding inference confidence. As shown below:

[0046] In some examples, when the inference confidence of a large language model... i exceeds the threshold Only when this condition is met will the data point be marked as a confirmed outlier.

[0047] Step 3: Dynamic calibration strategy selection and execution process. Based on the anomaly types identified in Step 2, an interpolation strategy is dynamically selected. For each identified anomaly, the data monitoring agent will retrieve data from the tool library. S The optimal correction strategy is selected, and the above process is implemented by the large language model by calling the appropriate tools. This decision is controlled by a multi-criteria utility function:

[0048] here, This represents the score of the improvement effect of strategy s in similar historical scenarios. This indicates its computational efficiency score. The weighting parameters are set as follows: , Prioritize the correction effect. The candidate strategies include linear interpolation, seasonal-trend interpolation based on STL decomposition, and substitution using the mean of similar spare parts. After a strategy is selected, the data intelligence system invokes the appropriate tools to perform the correction and subsequently updates the current data state, denoted as... .

[0049] Step 4: Normalization and Output: After completing all anomaly handling, the data monitoring AI will determine the current data status. Perform Min-Max normalization to map the values ​​to the [0, 1] interval, thereby generating the final cleaned dataset. The minimum and maximum values ​​used in this normalization will be saved to the global state. This allows for possible inverse transformations in subsequent steps.

[0050] In some embodiments, if the baseline quality score If the data quality is acceptable, the Step 2 deep anomaly detection step can be skipped, and the process can proceed directly to lightweight regularization, such as Z-score standardization.

[0051] Step 5: In some embodiments, the data monitoring phase also includes closed-loop feedback, where the data monitoring agent is designed to receive feedback. The feedback signal can be transmitted through the global state. The signal is transmitted. Upon receiving such a signal, the data monitoring agent can reassess its quality threshold. This mechanism, which uses anomaly detection parameters to determine whether to initiate a new cleaning cycle, enables the multi-agent intermittent demand prediction system to achieve adaptive optimization at the system level. Feedback signals can originate from other stages within the system, such as feature extraction, pattern classification, and model selection, to achieve dynamic path adjustment within the mechanism, significantly improving system fault tolerance.

[0052] In this embodiment, the data monitoring agent solves the problem of misjudgment in traditional cleaning methods when dealing with intermittent demands by integrating statistical detection and large model inference.

[0053] S2: As Figure 1 , Figure 5 and Figure 6 As shown, the feature extraction agent is the key transformation module from raw data to a predictive model. During the feature extraction stage, addressing the high-dimensional sparsity and heterogeneous distribution of spare parts demand sequences, the feature extraction agent is built upon a large language model. Using this large language model as the strategy engine, it dynamically analyzes the high-quality dataset through iterative reasoning. Based on the data characteristics, we independently plan the steps and strategies for feature engineering, and flexibly coordinate feature calculation and fusion tools to produce high-quality cleaned datasets. Transform into a highly sensitive feature set with high discriminative power and rich semantics. Specifically: Step 1: In some embodiments, the feature extraction stage may include a context-aware initialization and policy planning process. After the feature extraction agent is activated, it first queries the feedback signal in the global state, such as... Fields. If feedback exists from the model selection agent, the large language model will analyze the signal and use it as input to redesign the feature engineering strategy. For example, the feature extraction agent actively responds to external feedback from the pattern classification agent and the model selection agent. For instance, if the feedback signal indicates "insufficient lag features," the large language model can redesign, adding higher-order lag terms in the temporal feature extraction step and re-executing the entire refining process, forming a closed-loop optimization.

[0054] If there is no feedback, the default strategy is executed: first, the demand_pattern_classification() tool is called to classify the cleaned sparse demand sequence based on the SBC classification method. Perform initial pattern assessment, calculate the Average Demand Interval (ADI) and the squared coefficient of variation (CV) of demand. 2 They are categorized into smooth, intermittent, unstable, and blocky demands.

[0055] Step 2: In some embodiments, the feature extraction stage may include a parallel feature extraction tool scheduling process. Based on preliminary analysis or feedback instructions, the large language model formulates a detailed extraction plan and schedules three main types of feature extraction tools: (1) Statistical feature tools (calculate_statistical_features()): calculate window statistics. Set the sliding window to Every day, it calculates statistical characteristics such as mean, variance, skewness, and kurtosis within each window. Simultaneously, it calculates global intermittent indicators, such as the proportion of zero demand. Mean of non-zero demand and standard deviation (2) Temporal feature tool (extract_temporal_patterns()): Deep mining based on the tsfresh library. Key configurations include: (a) Extraction length is (a) The linear trend slope of the sliding window; (b) Perform a Fast Fourier Transform (FFT) and retain the previous... (c) The amplitude and phase of the main frequency components; (d) The intensity indices of the trend, seasonal and residual components are obtained by STL decomposition; (e) The autocorrelation coefficients with lags of 1, 7 and 30 are calculated; (3) Deep feature tool (generate_deep_features_autoencoder()): Train a three-layer fully connected autoencoder. The encoder structure is: input layer (84-dimensional) → ReLU (64-dimensional) → ReLU (32-dimensional) → latent features (16-dimensional). The mean squared error (MSE) is used as the reconstruction loss, and the training is carried out for 50 epochs with a batch size of 32. After training, the 16-dimensional vector output by the encoder is taken as the deep feature; After the above tools are executed in parallel, the generated features are concatenated into the original high-dimensional feature set. Its dimensions Up to several hundred.

[0056] like Figure 7 and Figure 8 As shown, Step 3: In some embodiments, the feature extraction stage may include an adaptive feature selection and fusion process, especially for high-dimensional features. The feature extraction agent then enters the feature refinement stage. The large language model first analyzes... Meta-features, such as mean variance and mean correlation between features, are obtained. Based on this analysis and downstream classification tasks, the large language model selects and executes a hierarchical feature selection strategy from a modular toolkit: ① Quick Filtering: Use the variance threshold filter in the select_features_RFE() tool to remove features with variance less than 1. The characteristics were obtained. .

[0057] ② Precise Selection: For classification tasks, the `select_features_SHAP()` tool is called first. A basic LightGBM model is fitted on the training set, and the importance ranking of the SHAP values ​​of each feature is calculated. The top-N features are retained. .

[0058] ③ Optional fusion: If there is high collinearity among features, LLM may decide to use principal component analysis (PCA) to fuse features, retaining 95% of the variance component, and finally outputting a low-dimensional, highly sensitive optimized feature set. .

[0059] Step 1: Input and Initialization Phase. The process first receives two key inputs: the raw, unprocessed feature set ( ) and specific downstream task types ( This will determine the direction of feature selection. Subsequently, the system performs an initialization operation, setting the current feature set to be processed as the original feature set, and simultaneously defining a method library containing various mature feature selection algorithms (such as Recursive Feature Emission (RFE), Variance Thresholding (VT), Random Forest (RF), SHAP Value, Lasso Regression, etc.).

[0060] Step 2: State Assessment and Intelligent Decision-Making Core. Before data processing, the system performs state assessment in parallel. It calculates the dimensionality and sparsity of the current feature set and deeply analyzes the task requirements, summarizing this key information into "output features" and passing them to the core "ReAct Decision (LLM Agent)" module. Utilizing the powerful reasoning capabilities of LLM, it analyzes the current state and, through the "action" step, selects the highest-scoring algorithm from the method library as the "best method" for the current situation. )".

[0061] Step 3: Hierarchical Filtering and Validation Feedback Loop. Feature processing employs a hierarchical strategy. First, in the "first stage," a fixed VT (variance threshold) algorithm is used for preliminary coarse filtering to obtain an intermediate feature set. Next, we move to the "second phase," applying the "best approach" selected by LLM to... Perform refined selection to generate candidate feature sets. Then comes the crucial "verification and feedback" stage, where a judgment is made. Does the quality meet the preset threshold? If the quality meets the standards, the process ends, and the final optimized feature set is output. If the quality does not meet the standards, a feedback mechanism is triggered, the "second-best method" is selected as a backup, and the process reverts to the second stage to re-filter until a satisfactory result is obtained.

[0062] like Figure 5 As shown, in some embodiments, the feature extraction agent includes a quality verification loop. During the generation... Then, it uses a simple benchmark model for a quick evaluation on the validation set. If the performance is lower than expected, LLM will re-evaluate the feature selection strategy.

[0063] It is understandable that the feature extraction agent follows the following... Figure 5The defined dynamic workflow is coordinated through a complete ReAct loop: LLM acts as the core inference and planning hub, first performing demand pattern analysis, then scheduling multiple feature computation tools in parallel, and finally executing adaptive feature selection and fusion. This mechanism, through adaptive input engineering strategies, fully leverages LLM's capabilities in context awareness and strategy planning, significantly improving the quality of feature representation for sparse and heterogeneous data, and overcoming the limitations of traditional methods in terms of insufficient feature expressiveness in such scenarios.

[0064] like Figure 9 As shown, after the feature extraction agent starts, the Large Language Model (LLM) first parses the intermittent demand data features and autonomously formulates a feature construction plan. Then, it calls specialized tools such as statistical feature calculation and temporal pattern mining according to the protocol. The LLM dynamically adjusts subsequent strategies based on the feedback from each tool output until a highly discriminative feature set is generated. For example, in the core reasoning stage of the ReAct loop (context-aware initialization and strategy planning), the LLM performs logical deduction based on the current data state. It does not directly handle low-level computation but analyzes data features and combines them with built-in knowledge to formulate the feature construction strategy for the current step. At this stage, the LLM analyzes and determines which mathematical methods can maximize the value of the data, thereby planning a specific action path, i.e., generating an intent plan. During the parallel feature extraction tool scheduling process, the LLM transforms the scheduled features into standardized execution instructions based on a defined strategy, which then interact with the standardized tool library on the right. The LLM precisely selects and calls functional modules from the tool library, such as `Calculate_statistical_features()` for basic statistics, `Extract_temporal_patterns()` for mining temporal patterns, or `Generate_deep_features()` for generating deep features. This process translates abstract semantic planning into concrete computational behavior. During the adaptive feature selection and fusion process, the computational results after tool execution are fed back to the LLM as observations in real time. The LLM receives and parses these feedback results, integrates them with previous contextual information, and updates the current cognitive state. This is a dynamically adjusted process: if the feedback indicates insufficient feature discrimination, the LLM will re-enter the inference stage based on the new state, adjusting the calling strategy or parameters. After multiple rounds of ReAct loops, when the LLM determines that the generated feature set has met the high discrimination requirement or reached the preset termination condition, the loop ends. The system outputs the final fused feature set.

[0065] S3: As Figure 1 , Figure 10 and Figure 11 As shown, the pattern classification agent uses the highly sensitive feature set output by the feature extraction agent. As input, its core objective is to transform numerical features into semantic category labels with clear business interpretability through cluster analysis and semantic inference. This provides crucial pattern-based information for the model's selection of agents. The pattern classification agent operates under the ReAct model, performing actions at each decision step. Its decision-making process can be formally represented as:

[0066] in, Indicates at time step The reasoning is based on the current state of the pattern classification agent. and historical experience The internal analysis and logical derivation process using the LLM large language model; Indicates at time step The actions taken were based on reasoning. The conclusion is that the pattern classification agent determines the specific operation to be performed; Indicates at time step The observations. This was made after the action was carried out. Subsequently, the pattern classification agent obtains feedback information or execution results from the environment or sub-agents.

[0067] Includes the current feature set Current clustering results The preliminary clustering results from the previous round (Initially empty), Comprehensive evaluation report of the previous round of clustering results Including indicators such as silhouette coefficient and Dunn index, while It is a record of historical clustering experiences that are similar to the current statistical characteristics.

[0068] like Figure 10 and Figure 11 As shown, the execution flow of the pattern classification agent follows a structured iterative loop, sequentially covering three stages: clustering strategy selection, execution and evaluation, and semantic generation. The specific steps are as follows: (1) In some embodiments, the pattern classification stage includes feature preprocessing and preliminary analysis, firstly processing the input high-sensitivity feature set. After standardization, quantitative diagnosis is performed, and meta-characteristics such as variance, quantiles, sparsity, and correlation density are calculated. .

[0069] (2) In some embodiments, the pattern classification stage includes reasoning-driven clustering strategy selection and iterative optimization: based on the initial state The ReAct loop is initiated. The large language model performs inference and selects the initial clustering algorithm. After execution, the result is... An evaluation report is generated during the observation phase. If the quality does not meet the standards, update the status to [new status]. LLM generation optimization suggestions This triggers a new cycle. This closed-loop mechanism of "evaluation-feedback-replanning" is the core of the agent's internal dynamic adaptation, enabling it to continuously optimize the algorithm and parameters based on real-time feedback of clustering results until a stable, high-quality clustering structure is obtained.

[0070] (3) In some embodiments, the pattern classification stage includes semantic label generation and structured output. After the ReAct loop terminates and stable clustering is obtained, the process enters the semantic mapping stage. For each finally determined cluster... The system will aggregate its key statistical features to form a comprehensive feature vector. LLM is based on Using a preset tag generation template, semantic generation is performed to conduct semantic reasoning and generation, producing outputs such as " High-Frequency, Stable, Growth-Consumable. "and other business semantic tags Ultimately, the pattern classification agent outputs a structured set of semantic labels. and Among them, feedback signal feedback_signal It can be applied to the data monitoring stage and / or feature extraction stage.

[0071] These semantically rich labels not only summarize current data patterns but also serve as a crucial hub driving the entire framework to achieve adaptive model selection. Different semantic categories directly and clearly indicate completely different demand behavior patterns, enabling the model selection agent to use these as a basis for differentiated and accurate optimal model retrieval and recommendation, achieving end-to-end adaptive linkage from data pattern recognition to prediction strategy formulation.

[0072] like Figure 12 As shown, the pattern classification agent begins its work with a deep understanding and strategy planning of the input highly sensitive feature set. Its built-in large language model calls feature statistical analysis tools through the ReAct architecture to optimize the feature set. Perform analysis and calculate the meta-feature vectors that reflect the characteristics of the data. (Including statistics such as mean, variance, and dimensionality). Based on this, the large language model performs logical reasoning based on meta-features and selects an initial clustering algorithm. This decision-making logic can be expressed as:

[0073] in, This represents a function that uses LLM for analysis and reasoning to generate strategy definitions. It refers to the optimal clustering algorithm determined after evaluation by large language model reasoning, covering typical methods such as K-Means, DBSCAN, and spectral clustering; This refers to the historical record library, used to store success cases.

[0074] When the initial algorithm The operation produces preliminary clustering results. Subsequently, the pattern classification agent initiates the clustering quality assessment and strategy optimization process: the system automatically calculates internal evaluation indicators such as the silhouette coefficient and the Calinski-Harabasz index, generating the optimal number of clusters K and the corresponding evaluation report. Then, the large language model integrates the report content with the original feature data for analysis, diagnoses the clustering effect, and generates optimization decision suggestions.

[0075] in, This represents a function defined for evaluating clustering results using LLM inference. If the evaluation... If the result is "True", the system will proceed to the next processing stage; if it fails, the pattern classification agent will optimize accordingly. Adjust the clustering strategy and re-execute the clustering to form a self-optimizing decision-making loop.

[0076] Once the clustering results reach a stable state, each cluster... It will be transformed into a comprehensive statistical feature vector This vector covers the average demand interval. Demand variation coefficient squared Average non-zero demand Zero demand ratio Trend strength These indicators are multi-dimensional and include statistical features such as quartiles, mean, and variance. The core task of LLM is to semantically parse these multi-dimensional statistical features and map them into intuitively understandable business category labels. This semantic generation process can be formally expressed as:

[0077] in, This represents a function that uses LLM for analysis and reasoning to generate semantic labels for each cluster. Indicates the first Semantic tags corresponding to cluster requirements K represents the optimal number of clusters generated by clustering. Specifically, if the statistical characteristics of a certain cluster are as follows: Its demand pattern can be summarized as "short intervals, small fluctuations, stable demand, and a weak growth trend." After comprehensively analyzing these characteristics, LLM will generate a semantic label for it as "high-frequency, stable-growth consumable." Finally, the pattern classification agent outputs a structured set of semantic labels. Each label originates from a deep analysis of multidimensional data patterns and includes key feature descriptions, providing a traceable and rich semantic basis for the subsequent model's decision to select agents.

[0078] S4: As Figure 1 and Figure 13 As shown, the prediction model selects an agent as the decision-making terminal for the multi-agent intermittent demand prediction method and system, responsible for integrating the upstream processing results—the highly sensitive feature set. and semantic tag set By integrating an external knowledge base through a Retrieval Enhancement Generation (RAG) mechanism, the system recommends the optimal prediction model for each demand pattern. Simultaneously, as the core feedback source of the system, it transforms the evaluation results of downstream prediction performance into structured instructions, driving the optimization of the feature extraction agent and the classification agent, forming a globally adaptive closed loop.

[0079] In some embodiments, the overall execution flow of the model selecting the agent is an iterative loop that includes knowledge retrieval, reasoning ranking, and verification feedback. The specific implementation steps can be as follows: Step 1: The model selection stage can include meta-feature construction and knowledge retrieval processes. The model selection agent process begins with the construction of a comprehensive meta-feature vector. This vector fused Statistical, time series, and deep features And category confidence. Based on meta-features. The model selects agents using the RAG mechanism as the front end for knowledge focusing, drawing from an external structured model semantic library. It performs precise semantic retrieval to recall the Top-K most relevant subset of candidate models. This step demonstrates the agent's ability to dynamically access and utilize external knowledge based on the current task context, avoiding the limitations of outdated or insufficient knowledge within large language models, and achieving adaptive expansion of recommendation sources.

[0080] Step 2: Inference-driven multi-criteria model analysis and ranking: Obtaining a subset of candidate models Then, the model selects an agent to enter the core inference loop. The large language model receives a detailed semantic description of each candidate model, including its principles, applicable scenarios, parameter requirements, historical performance examples, current meta-features, and semantic labels. Based on this, the large language model performs deep matching and multi-criteria joint inference. Finally, the large language model outputs an interpretably ranked list of models. Each model is then accompanied by a recommendation reason. This process deeply integrates data features, business semantics, and model knowledge, achieving an intelligent leap from feature matching to semantic-principle matching.

[0081] Step 3: Validation, Error Attribution and Feedback Generation: Perform rigorous cross-validation on the top-ranked models, calculate key indicators such as Mean Absolute Error (MAE), Mean Absolute Scaled Error (MASE), and MIS, and conduct error attribution analysis based on the large language model to identify the weaknesses of candidate models in specific scenarios.

[0082] If the performance of all candidate models fails to meet the threshold requirement, or if the error pattern points to a systematic upstream defect, the model selection agent will initiate a feedback optimization mechanism. It will generate a structured, actionable diagnostic report that clearly identifies the root cause of the problem and provides optimization recommendations. This report is written to the global state via a strict JSON interaction protocol. Feedback signal This field. Upon detecting this signal, the state graph engine can trigger a backtracking of the workflow within the agent through conditional edges. For example, it can reactivate the feature extraction agent and inject the optimization instruction, thereby initiating a new, targeted feature engineering optimization cycle. This mechanism allows the model selection agent to transcend the role of a single recommender, becoming the core engine driving the entire multi-agent intermittent demand prediction system to achieve global, closed-loop, performance-oriented adaptive behavior.

[0083] Step 4: Final Recommendation Generation and Knowledge Base Evolution: After obtaining satisfactory verification results, the results output module will generate the final, interpretable recommendation report. Includes the best model The document includes confidence levels, detailed justifications, parameter configuration recommendations, and risk warnings. It also covers the meta-features of this prediction task. Final Recommendation Model Validation errors and contextual information will be stored as a new knowledge unit in the external model semantic knowledge base. This enables the system to continuously enrich its decision-making basis through continuous learning, achieving long-term adaptive capability where the recommendation strategy evolves with the accumulation of experience.

[0084] like Figure 14 As shown, in some embodiments, to avoid burdening large language models with lengthy original sparse sequences of requirements and high-dimensional features directly input, the feature extraction agent and the pattern classification agent adopt a unified strategy: first, clean the data at the tool layer. Perform in-depth statistical analysis to extract refined meta-features that characterize the essence of the data. The large language model is used as input for semantic summarization. For example, the large language model in the feature extraction agent plans feature engineering paths based on meta-features, while the pattern classification agent selects clustering algorithms based on meta-features. This design allows the large language model to focus its attention on the macroscopic statistical properties that determine data patterns, significantly reducing processing complexity.

[0085] The selection of agents by the model needs to be done from the candidate model library. We recommend suitable prediction models. To address the knowledge overload problem, we introduce Retrieval Augmentation (RAG) technology as a dynamic knowledge filtering front-end, combined with the meta-features output by the pattern classification agent. With semantic tags To construct a model recommendation mechanism for intermittent demand prediction: First, semantic retrieval is used to recall the most relevant Top-K candidate model subsets from the knowledge base. :

[0086] By using the RAG mechanism to filter a subset of Top-K model candidates, the open-domain model selection problem is essentially converged to a finite and highly relevant decision space. Subsequently, the large language model does not need to process the complete knowledge base, but only needs to be based on the structured candidate model descriptions. The model performs deep comparative reasoning based on its contextual information, ultimately generating the model's recommendation results.

[0087] This refers to the optimal prediction model ultimately selected for each type of feature. This mechanism ensures that the LLM inference process is always based on refined knowledge units that are highly relevant to the current task, thereby significantly improving the accuracy of model recommendations and the interpretability of decisions.

[0088] These strategies form a unified system within the framework: they are essentially task-context-oriented information preprocessors. Their adaptive nature is reflected in the fact that the input content is not a fixed template, but dynamically generated based on the upstream processing results and the current task state. Through end-to-end input engineering design, we systematically transform the original, complex, sparse, and variable industrial data streams into refined task briefings that conform to the cognitive characteristics of LLM, thereby ensuring the comprehensive performance of LLM-based multi-agent collaborative processes in terms of robustness, execution efficiency, and decision transparency.

[0089] The model selects the agent's complete decision-making process, which follows a strict ReAct paradigm. At each decision step... The ReAct loop for model agent selection can be formally represented as:

[0090] in, Indicates at time step The reasoning behind this is that the model selects the agent based on the current state. and historical experience The internal analysis and logical derivation process using LLM; Indicates at time step The actions taken were based on reasoning. The conclusion is that the model selects the agent to determine the specific operation to be performed; Indicates at time step The observations. This was made after the action was carried out. Then, the agent receives feedback information or execution results from the environment or sub-agents.

[0091] System status Includes highly sensitive feature sets With semantic tag set Through the RAG mechanism, semantic knowledge bases from external models are obtained. The subset of Top-K candidate models most relevant to the current task, retrieved from the database, along with the performance metrics, error attribution analysis results, and successful model recommendation records under similarity features.

[0092] like Figure 15 As shown in some embodiments, at the level of inter-agent communication, to avoid semantic distortion and formatting issues common in multi-turn dialogues, the framework constructed in this application abandons free-format natural language and mandates the use of a JSON-based structured interaction protocol. This protocol stipulates that all agent outputs must include three parts: metadata, core data payload, and a record of the inference process.

[0093] The information transmitted from the pattern classification agent to the model selection agent follows this standardized format. This design ensures that downstream agents can accurately parse numerical task data and enhance decision-making logic through semantic reasoning trajectories, achieving a balance between data accuracy and semantic consistency. For example, the information transmitted from the pattern classification agent to the model selection agent first interprets the received pattern features. By reading semantic labels and analyzing key feature vectors, it comprehensively judges that the sequence transmitted by the pattern classification agent is a high-frequency, intermittent, low-fluctuation, and weak-trend demand pattern. Then, it matches the pattern selection constraints, recommends suitable models according to priority, and clarifies the models that are not recommended and the reasons for them.

[0094] In some embodiments, this application builds an experimental platform based on the DeepSeek-V3 API, connects to the official interface, and uses the LangGraph framework to construct an agent collaborative system. This model, with its excellent reasoning, contextual understanding, and tool invocation capabilities, becomes the unified decision-making core for all agents in the system. Regarding parameter settings, the temperature parameter can be adjusted to 0.2 to reduce the randomness in the feature generation and policy planning stages, ensuring stable and reliable output results.

[0095] The coordinated scheduling of the entire system is implemented using the LangGraph framework. Its directed graph structure clearly defines the state transitions and interaction dependencies between sub-agents, thereby realizing the controllable complex task pipeline described in the algorithm flow. The system is deployed on a local computing node, and the hardware uses a 13th generation Intel Core i5-13600KF processor. This configuration can efficiently complete agent scheduling, business logic processing, and local feature calculation in a resource-constrained environment, ensuring the overall operating efficiency and stability of the system.

[0096] In some embodiments, this application selects three publicly available industry datasets with significant differences to systematically verify the comprehensive performance of the multi-agent framework in intermittent demand forecasting. The selection criteria are as follows: (1) Industry representativeness: covering aviation, automobile manufacturing and general manufacturing, corresponding to spare parts demand scenarios with different supply chain characteristics and management models.

[0097] (2) Pattern coverage: The three datasets together cover the spectrum of intermittent demand from extremely sparse to relatively continuous, which can comprehensively test the generalization and adaptability of the framework.

[0098] (3) Universality: The selected datasets are all benchmark data widely used in the field, ensuring that the experimental results are comparable and easy to compare with existing methods.

[0099] The datasets are briefly described below: Dataset 1: Intermittent Aircraft Spare Parts (RAF) Demand: Contains monthly demand for 5,000 types of aircraft spare parts over 84 months. Aircraft spare parts are characterized by high value, long lifespan, and extremely sparse demand. Their extremely high average demand interval (ADI) and coefficient of variation squared (CV²) are typical samples for testing the model's ability to handle highly sparse scenarios.

[0100] Dataset 2: US Automotive Parts Demand: Covers monthly demand for 2,674 parts over 51 months. 1,992 valid sequences were retained after cleaning. Automotive supply chain demand is relatively stable. The mean demand interval (ADI) and coefficient of variation (CV²) of this dataset are significantly lower than those of aviation data, making it suitable for validating the model's performance in scenarios transitioning from intermittent to continuous demand.

[0101] Dataset 3: Spare Parts Sales in the Dutch Manufacturing Industry: Contains weekly sales data for 3,451 products over 150 weeks. Demand for spare parts in the manufacturing industry is influenced by production planning and maintenance cycles, exhibiting moderate intermittency and concentrated bursts. Its demand scale and volatility are widely distributed, making it suitable for testing the framework's scalability in handling highly heterogeneous SKU pools.

[0102] To clearly present the data characteristics, Tables 2 and 3 summarize the non-zero demand mean for each dataset ( Size_mean ), standard deviation ( Size_std Core statistics include the coefficient of variation squared (CV²) and the average demand interval (ADI).

[0103] Table 2 Core statistics of dataset 1 distributed Size_mean Size_std CV2 ADI Minimum value 1.00 0.00 0.00 3.47 Upper quartiles 1.56 0.81 0.18 7.40 median 3.83 3.06 0.46 9.38 lower quartiles 11.33 9.35 0.90 11.14 Maximum value 668.00 874.42 11.88 26.67 Table 3 shows the core statistics of dataset 2. distributed Size_mean Size_std CV2 ADI Minimum value 1 0 0 1.34 Upper quartiles 1.38 0.55 0.14 2.43 median 1.67 0.92 0.28 3.64 lower quartiles 2.12 1.36 0.43 6.38 Maximum value 15 21.14 4.32 10.2 Table 4 shows the core statistics of dataset 3. distributed Size_mean Size_std CV2 ADI Minimum value 0.31 0 0 1.01 Upper quartiles 3.66 2.97 0.42 3.85 median 8.99 7.84 0.78 7.89 lower quartiles 22.91 21.76 1.36 15 Maximum value 1132.21 853.5 14.99 25 A comprehensive comparison of the above statistical characteristics reveals the following: In terms of demand scale distribution, both the RAF and Dutch datasets exhibit significant long-tail characteristics, with a small number of highly heterogeneous SKUs dominating the demand, while the US automotive dataset shows a relatively uniform distribution. Regarding intermittency and volatility, the RAF dataset has the longest demand intervals and the greatest volatility, representing an extremely intermittent type; the Dutch dataset falls in the middle; and the US automotive dataset is the most stable. These three datasets together constitute a complete demand spectrum ranging from stable and continuous to highly intermittent, providing a sufficient data foundation for system validation of the prediction accuracy, interpretability, and stability of the multi-agent framework in different scenarios.

[0104] In some embodiments, multiple representative prediction methods are selected as baselines for comparison to systematically test the comprehensive performance of the multi-agent framework. This includes classical statistical models, cutting-edge machine learning and meta-learning methods, as well as emerging LLM large language prediction technology, ensuring the comprehensiveness and hierarchy of the evaluation system.

[0105] (1) Classical statistical methods As a benchmark method in the field of intermittent demand forecasting, such models are known for their simplicity and stability in handling sparse data, including the Croxton basic method and its classic improved models SBA and TSB.

[0106] 1) The Croston Method The core idea of ​​this method is to decompose the demand sequence into two subsequences: demand interval and demand quantity, and then perform exponential smoothing on each. The basic calculation formula is as follows: If :

[0107] If :

[0108] In the formula, This is an estimate of the demand. This is an estimate of the demand interval.

[0109]

[0110] This application uses the Croston method as the most fundamental comparison model in the benchmark system. Its smoothing parameters... Adaptive optimization is performed using the Nelder-Mead optimization algorithm, with the objective function being to minimize the in-sample mean squared error. MSE ).

[0111] The parameter optimization process can be summarized as follows:

[0112] 2) Syntetos-Boylan Modified Model (SBA) This model, as an improved version of the Croston method, enhances prediction accuracy by introducing a bias correction factor. Its core correction formula is as follows:

[0113] The SBA model has been widely adopted in practice. The smoothing parameters of this model... The parameters are adaptively determined using the maximum likelihood estimation method, which assumes that the prediction error follows a normal distribution. The optimal parameters are obtained by maximizing the following log-likelihood function:

[0114] 3) Teunter-Syntetos-Babai (TSB) model The TSB model achieves a significant improvement by optimizing the probability update mechanism during periods of zero demand. This feature allows it to better adapt to recovery scenarios following sudden demand disruptions. Its demand occurrence probability... The update rules are as follows: when observed hour:

[0115] In the formula The smoothing coefficient is adaptively selected using a time-series cross-validation method, with the optimization objective being to minimize the mean square error.

[0116]

[0117] (2) Machine learning methods Machine learning methods excel at capturing non-linear features of data, but unlike the strategy of extracting meta-features from the overall sequence in the embodiments of this application, these methods are typically based on sliding windows to extract local temporal features. The experiments selected three types of models commonly used in the field of intermittent demand forecasting: Random Forest (RF), Long Short-Term Memory (LSTM) networks, and Lightweight Gradient Boosting Machine (LightGBM).

[0118] 1) Random Forest (RF) Based on the Bagging ensemble framework, multiple decision trees are constructed through feature randomness and bootstrap sampling to reduce variance. The input features include the statistical meta-features of each SKU and the lag values ​​of the first 1-3 periods. Key hyperparameters are adaptively selected within a set range through Bayesian optimization (number of decision trees 100-500, maximum depth 5-10 or unlimited, minimum number of leaf nodes 1-5). The optimization objective is to minimize the mean squared error of 5-fold cross-validation.

[0119] 2) Lightweight Gradient Booster (LightGBM) Based on the gradient boosting decision tree framework, a histogram algorithm and leaf growth strategy are used to improve efficiency. The input features are the same as those of random forest. Bayesian optimization is used to adjust parameters such as the number of leaf nodes (31-127), learning rate (0.05-0.2), and feature sampling ratio (0.7-1.0), with the objective of minimizing the 5-fold cross-validation error.

[0120] 3) Long Short-Term Memory Network (LSTM) A gating mechanism is used to capture long-term dependencies in the demand sequence. The Adam optimizer is used for training, and the hidden layer dimension (50-150) and learning rate (1e-4 to 1e-2) are adaptively adjusted using the Hyperband early stopping method. Training is stopped when the loss function shows no improvement for 5 consecutive rounds.

[0121] (3) Meta-learning and ensemble methods Such methods aim to automate the model selection process, consistent with the model selection agent objective proposed in this paper, but their implementation mechanism is to assign weights to candidate models through time series features.

[0122] 1) Feature-Driven Prediction Model Averaging (FFORMA) XGBoost is used as the meta-model to learn the mapping relationship between sequence meta-features and the prediction performance of candidate models (the three statistical methods and three machine learning models mentioned above). The meta-model hyperparameters are adaptively optimized through cross-validation, with the objective of minimizing the MSE loss.

[0123] 2) Weighted Average Fusion Strategy (WAF) An ensemble prediction is generated by weighted summation of predictions from multiple candidate models. The weight combination is determined by minimizing the mean square error of the ensemble prediction on the validation set, and the optimal configuration is obtained by traversing the weight space using a grid search.

[0124] (4) Large Language Model Method To verify the advantages of the "LLM as a coordinating center" paradigm over the direct prediction paradigm, we introduce the following baseline method: 1) Chronos Model (Large-Scale Time Series Forecasting Model) Based on the Transformer architecture, continuous temporal values ​​are discretized into lexical units for language pattern prediction. Zero-shot prediction is performed directly using the pre-trained Chronos-T5-large model provided by Hugging Face to evaluate the cross-domain adaptability of the large language model to intermittent data.

[0125] 2) LLM zero-shot hint method Time-series data is converted into text input to the DeepSeek-R1 model, which leverages its semantic understanding capabilities to generate predictions. The temperature parameter is set to 0 to reduce randomness, and a fixed prompt template is used: "Based on the past 72 demand sequences (including zero values), predict the demand value for the next period. Only the number is returned without explanation." Each sequence is predicted three times independently, and the median is taken as the final result.

[0126] By comparing the system with the aforementioned baseline methods, the comprehensive performance advantages of the multi-agent framework in terms of prediction accuracy, stability, and interpretability can be fully verified.

[0127] In some embodiments, this application provides a systematic evaluation of the performance of the provided framework. To comprehensively examine its performance on datasets with varying degrees of intermittency, we construct a multi-dimensional evaluation system: employing the mean absolute error (MAO)... MAE ) and root mean square error ( RMSE To measure the accuracy of point prediction, the mean absolute ratio error (MAR) is introduced. MASE (Relative Mean Absolute Error) RelMAE To eliminate the effects of data scaling differences and zero inflation, average interval scores were used. MIS Assess the reliability of the prediction interval and use pairing t Test the statistical significance of the performance differences.

[0128] This application systematically compares the proposed method with multiple baseline methods on three heterogeneous datasets, comprehensively examining the predictive capabilities of the multi-agent framework provided in the embodiments of this application under different intermittent patterns and industry scenarios. The evaluation assesses the prediction accuracy of content coverage points, intermittent data-specific metrics, and prediction interval quality. To verify the statistical significance of performance differences, paired methods are used. tThe test performs statistical inference on performance differences arising from the same set of time-series data. Optimal results for each evaluation metric are highlighted in bold, and statistical significance markers are set as follows: , , and .

[0129] For detailed assessment results, please refer to Tables 5 to 7 below: Table 5 Evaluation of prediction results for Dataset 1 Predictive Model MAE RMSE MASE RelMAE MIS paired t-test(p-value) Croston 8.45±0.31 15.21±0.68 1.32±0.05 1.28±0.04 12.35±0.45 <0.001 *** SBA 8.21±0.29 14.89±0.65 1.29±0.05 1.25±0.04 12.01±0.43 <0.001 *** TSB 8.05±0.28 14.70±0.63 1.26±0.04 1.22±0.04 11.85±0.41 <0.001 *** Random Forest 7.92±0.27 14.55±0.61 1.24±0.04 1.20±0.04 11.72±0.40 0.002 ** LightGBM 7.71±0.25 14.32±0.58 1.21±0.04 1.17±0.03 11.50±0.38 0.012 * LSTM 7.85±0.26 14.48±0.60 1.23±0.04 1.19±0.04 11.65±0.39 0.005 ** Chronos 7.35±0.23 13.65±0.52 1.15±0.03 1.11±0.03 10.95±0.34 0.038 * LLM zero-sample 7.42±0.24 13.78±0.54 1.16±0.03 1.12±0.03 11.08±0.35 0.031 * FFORMA 7.38±0.23 13.70±0.53 1.15±0.03 1.11±0.03 11.02±0.34 0.035 * WAF 7.45±0.24 13.82±0.54 1.16±0.03 1.12±0.03 11.12±0.35 0.029 * LIMA 6.98±0.21 12.85±0.48 1.10±0.03 1.06±0.02 10.54±0.30 (Benchmark) Table 6 Evaluation of prediction results for Dataset 2 Predictive Model MAE RMSE MASE RelMAE MIS Paired t-test (p-value) Croston 2.15±0.08 3.85±0.15 1.15±0.03 1.12±0.03 4.21±0.18 <0.001 *** SBA 2.09±0.07 3.76±0.14 1.12±0.03 1.09±0.03 4.08±0.17 <0.001 *** TSB 2.04±0.07 3.69±0.14 1.10±0.03 1.07±0.03 3.98±0.16 <0.001 *** Random Forest 1.98±0.06 3.61±0.13 1.08±0.03 1.05±0.03 3.85±0.15 0.003 ** LightGBM 1.92±0.06 3.52±0.12 1.05±0.02 1.02±0.02 3.72±0.14 0.015 * LSTM 1.95±0.06 3.58±0.13 1.07±0.03 1.04±0.03 3.80±0.15 0.007 ** Chronos 1.80±0.05 3.30±0.11 0.98±0.02 0.95±0.02 3.45±0.12 0.042 * LLM zero-sample 1.82±0.05 3.33±0.11 0.99±0.02 0.96±0.02 3.48±0.13 0.039 * FFORMA 1.78±0.05 3.28±0.10 0.97±0.02 0.94±0.02 3.42±0.12 0.045 * WAF 1.81±0.05 3.32±0.11 0.98±0.02 0.95±0.02 3.46±0.12 0.041 * LIMA 1.72±0.04 3.18±0.10 0.95±0.02 0.92±0.02 3.32±0.11 (Benchmark) Table 7 Evaluation of prediction results for Dataset 3 Predictive Model MAE RMSE MASE RelMAE MIS Paired t-test (p-value) Croston 5.82±0.22 10.45±0.45 1.25±0.04 1.21±0.04 9.12±0.35 <0.001 *** SBA 5.68±0.21 10.22±0.43 1.22±0.04 1.18±0.04 8.89±0.33 <0.001 *** TSB 5.58±0.20 10.05±0.41 1.20±0.04 1.16±0.04 8.72±0.32 <0.001 *** Random Forest 5.45±0.19 9.88±0.40 1.18±0.04 1.14±0.04 8.55±0.31 0.004 ** LightGBM 5.32±0.18 9.65±0.38 1.15±0.03 1.11±0.03 8.32±0.29 0.018 * LSTM 5.40±0.19 9.78±0.39 1.17±0.04 1.13±0.04 8.48±0.30 0.009 ** Chronos 5.10±0.17 9.15±0.35 1.10±0.03 1.06±0.03 7.90±0.27 0.048 * LLM zero-sample 5.15±0.17 9.25±0.36 1.11±0.03 1.07±0.03 8.00±0.28 0.044 * FFORMA 5.05±0.16 9.08±0.34 1.09±0.03 1.05±0.03 7.85±0.26 0.052 * WAF 5.12±0.17 9.20±0.35 1.10±0.03 1.06±0.03 7.95±0.27 0.046 * LIMA 4.85±0.15 8.75±0.32 1.04±0.03 1.00±0.02 7.61±0.25 (Benchmark) Based on experimental results from three datasets, the performance of the system framework provided in this application embodiment was analyzed in depth. The results demonstrate that in highly intermittent scenarios, such as the RAF dataset, this application embodiment effectively overcomes the challenges posed by extremely sparse data through a multi-agent collaboration mechanism—especially the data quality monitoring and dynamic model selection module.

[0130] In a dataset of US auto parts with relatively stable demand, the multi-agent intermittent demand forecasting system provided in this application has a significantly lower error than a single machine learning model, demonstrating the value of model integration and external knowledge fusion. It maintains its leading position by virtue of its comprehensive index advantages, and statistical test results confirm the significance of its advantages.

[0131] In the Dutch manufacturing dataset with mixed demand patterns, the multi-agent intermittent demand forecasting system provided in this application continues to demonstrate stable performance advantages, achieving optimal balance across various indicators through end-to-end agent collaboration.

[0132] Cross-dataset experiments consistently demonstrate that the multi-agent intermittent demand forecasting system provided in this application can effectively adapt to demand scenarios with varying degrees of intermittency. From extremely intermittent to relatively stable modes, the multi-agent intermittent demand forecasting system provided in this application exhibits a sustained performance advantage, thanks to the collaborative design of its multi-agent architecture: data quality monitoring ensures input cleanliness, feature extraction and classification deepen the understanding of demand patterns, and intelligent model selection achieves optimal matching between strategies and data. Statistical tests further verify the reliability of the conclusions. This framework provides a practical solution for spare parts demand forecasting in the field of intelligent manufacturing, enabling accurate and reliable forecasting performance in complex real-world scenarios.

[0133] like Figure 15As shown, in some embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the numerical simulation, analysis, or control steps in any of the methods described above.

[0134] At the hardware level, this electronic device includes a processor, internal bus, network interface, memory, and non-volatile storage, and may also include other hardware required for the business logic. The processor reads the corresponding computer program from the non-volatile storage into memory and then executes it to achieve the above-mentioned functions. Figure 1 This specification does not exclude other implementation methods besides software implementation, such as logic devices or a combination of hardware and software. In other words, the execution entity of the following processing flow is not limited to individual logic units; it can also be hardware or logic devices. It is understood that by simply performing some logic programming on the method flow using a hardware description language and programming it into an integrated circuit, the hardware circuit implementing the logic method flow can be obtained.

[0135] The above method can be implemented by a controller in any suitable manner. For example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, microcontrollers. A memory controller can also be implemented as part of the control logic of a memory. It is understood that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, ASICs, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Or even, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0136] The systems, devices, modules, or units in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0137] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this specification, the functions of each module can be implemented in one or more software and / or hardware.

[0138] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0140] In some embodiments, this application may also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the prediction method described above.

[0141] In some embodiments, these computer program instructions may also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0142] In some embodiments, these computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0143] In some embodiments, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0144] In some embodiments, memory may include non-persistent memory in a computer-readable medium, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0145] In some embodiments, computer-readable media, including permanent and non-permanent, removable and non-removable media, can be used to store information by any method or technology. The information can be computer-readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0146] Those skilled in the art will readily understand that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An LLM-enabled multi-agent intermittent demand forecasting method, characterized in that, It includes the following stages: data monitoring, feature extraction, pattern classification, and model selection; During the data monitoring phase, the data monitoring agent relies on the semantic parsing and context discrimination capabilities of the large language model to intelligently schedule data cleaning tools to transform the input raw intermittent demand sequence containing mixed noise into a high-quality dataset. In the feature extraction stage, the feature extraction agent uses the large language model as its strategy brain, dynamically analyzes the data characteristics of the high-quality dataset through iterative reasoning, autonomously plans the steps and strategies of feature engineering, and coordinates feature computing and fusion tools to transform the high-quality dataset into feature combinations that can capture complex patterns. In the pattern classification stage, the pattern classification agent relies on the strategy selection and semantic generation function of the large language model. First, it matches a suitable clustering algorithm according to the feature structure of the feature combination, and then generates readable labels based on the statistical characteristics of the clustering algorithm results, transforming the feature combination into a requirement pattern category with business semantics. In the model selection phase, the model selection agent uses the large language model as a reasoning engine that integrates external knowledge. It calls the structured model knowledge base through retrieval enhancement generation technology, combines data features with the demand pattern categories to perform multi-dimensional inference, and finally outputs a model recommendation scheme that matches each type of demand pattern.

2. The LLM-enabled multi-agent intermittent demand forecasting method according to claim 1, characterized in that, Prior to the data monitoring phase, the prediction method further includes establishing a modular shared tool library, which is invoked during the data monitoring phase, the feature extraction phase, the pattern classification phase, and the model selection phase, and includes the tools required for the data monitoring phase, the feature extraction phase, the pattern classification phase, and the model selection phase.

3. The LLM-enabled multi-agent intermittent demand forecasting method according to claim 2, characterized in that, The modular shared toolkit includes an anomaly monitoring and handling module, a data preprocessing module, a data standardization module, a statistical feature calculation module, a time series feature calculation module, a TSfresh time series feature calculation module, SBC classification, an autoencoder feature extraction module, recursive feature elimination, SHAP feature importance analysis, K-means clustering, hierarchical clustering, cluster evaluation, historical storage, LightGBM model prediction, TBS model prediction, RAG TOP-K model selection, and operation report generation.

4. The LLM-enabled multi-agent intermittent demand forecasting method according to claim 2 or 3, characterized in that, Prior to the data monitoring step, the prediction method further includes establishing a standardized invocation protocol. The standardized invocation protocol is used to construct a dynamic closed-loop mechanism for perception, decision-making, and execution. The data monitoring agent, the feature extraction agent, the pattern classification agent, and the model selection agent follow the standardized invocation protocol when invoking the modular shared tool library. Furthermore, the data monitoring agent, the feature extraction agent, the pattern classification agent, and the model selection agent form a whole through the standardized invocation protocol.

5. The LLM-enabled multi-agent intermittent demand forecasting method according to claim 4, characterized in that, The scheduling of the multi-agent collaboration process within the prediction method is handled by a state graph execution engine built on the LangGraph framework. The engine positions the large language model as the intelligent decision-making center. The multi-agent includes the data monitoring agent, the feature extraction agent, the pattern classification agent, and the model selection agent.

6. The LLM-enabled multi-agent intermittent demand forecasting method according to claim 5, characterized in that, The multi-agent collaboration process in the prediction method is a directed state graph. The engine has a global short-term memory. In the data monitoring stage, the feature extraction stage, the pattern classification stage, and the model selection stage, the multi-agent writes processing data and running status to the global short-term memory in real time to achieve information sharing and backtracking.

7. The LLM-enabled multi-agent intermittent demand forecasting method according to claim 6, characterized in that, When the modular shared tool library is invoked by the data monitoring agent, the feature extraction agent, the pattern classification agent, and the model selection agent, the large language model selects a specific tool through logical reasoning, generates standardized instructions to drive the modular shared tool library to execute, and updates the status based on the execution results to enter the next round of decision-making.

8. The LLM-enabled multi-agent intermittent demand forecasting method according to claim 7, characterized in that, The prediction method includes an adaptive feedback integration mechanism, which supports the dynamic path adjustment of the multi-agent collaborative process of the prediction method. When the prediction error exceeds the limit in the model selection step, the engine rolls back the collaborative process state to the feature extraction node through the conditional transition path. If the received signal indicates that the lagging features are insufficient, the large language model is replanned, and the adaptive feedback integration mechanism automatically generates dynamic instructions to guide the feature extraction to add short-term lagging features.

9. An LLM-enabled multi-agent intermittent demand forecasting system, characterized in that, This includes data monitoring agents, feature extraction agents, pattern classification agents, and model selection agents; The data monitoring agent is used to perform the semantic parsing and context discrimination capabilities based on the large language model as described in any one of claims 1-8, and to intelligently schedule data cleaning tools to transform the input raw intermittent demand sequence containing mixed noise into a high-quality dataset. The feature extraction agent is used to perform the process described in any one of claims 1-8, which uses a large language model as the strategy brain, dynamically analyzes the data characteristics of a high-quality dataset through iterative reasoning, autonomously plans the steps and strategies of feature engineering, and flexibly coordinates feature computing and fusion tools to transform a high-quality dataset into a feature combination that can capture complex patterns. The pattern classification agent is used to perform the strategy selection and semantic generation function based on the large language model as described in any one of claims 1-8. First, it matches a suitable clustering algorithm according to the feature structure of the feature combination, and then generates readable labels based on the statistical characteristics of the clustering algorithm results, thus transforming the feature combination into a requirement pattern category with business semantics. The model selection agent is used to execute the process described in any one of claims 1-8, which uses the large language model as a reasoning engine that integrates external knowledge, calls the structured model knowledge base through retrieval enhancement generation technology, combines data features with the demand pattern categories to perform multi-dimensional inference, and finally outputs a model recommendation scheme that matches each type of demand pattern.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the prediction method as described in any one of claims 1-8.

Images