System, apparatus, and method for time-series forecasting

The neural network based method effectively predicts future events using neural networks.

WO2026235421A1PCT designated stage Publication Date: 2026-11-12LG MANAGEMENT DEV INST CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/KR2026/095991
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2026-03-27
Filing Date
2026-04-16
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing methods for predicting future events using neural networks fail to provide a robust and efficient way of predicting future events using neural networks.

Method used

A neural network based method for predicting future events using neural networks.

Benefits of technology

The method provides a robust and efficient way to predict future events using neural networks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure KR2026095991_12112026_PF_FP_ABST
    Figure KR2026095991_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A prediction data generation method according to one embodiment of the present disclosure may comprise the steps of: acquiring time-series data; generating a plurality of pieces of randomly extracted data by performing random extraction a plurality of times on the time-series data; performing a time-axis attention operation that models the temporal dependency of the plurality of pieces of randomly extracted data; performing an inter-feature attention operation that extracts the correlation between a plurality of features of the plurality of pieces of randomly extracted data; and generating at least one piece of prediction data on the basis of a result of the time-axis attention operation and a result of the inter-feature attention operation.
Need to check novelty before this filing date? Find Prior Art

Description

Time series forecasting system, device and method

[0001] The present disclosure relates to a time-series forecasting system, apparatus, and method, and more specifically, to a system, apparatus, and method for predicting the future using data sequentially recorded over time using a neural network.

[0002] Time series data refers to data recorded sequentially over time. The problem of predicting the future by analyzing observed time series data is the time series forecasting problem. Recently, research on predicting the future using multivariate time series data has been primarily conducted. Since these variables can range from hundreds to millions, algorithms capable of efficiently processing and learning them are becoming very important. The present disclosure aims to provide such an efficient and accurate multivariate time series forecasting system, apparatus, and method.

[0003] The present disclosure aims to provide an efficient and highly accurate multivariate time series forecasting system, apparatus, and method.

[0004] A system according to one embodiment of the present disclosure comprises one or more processors; and one or more memories that collectively store instructions that cause the system to perform operations when executed by the one or more processors, wherein the operations may include: an operation of acquiring time series data; an operation of generating a plurality of randomly extracted data by performing a plurality of random extractions on the time series data; an operation of performing an operation of modeling the temporal dependency of the plurality of randomly extracted data; an operation of performing an operation of extracting correlations between a plurality of features of the plurality of randomly extracted data; and an operation of generating at least one prediction data based on at least one of the result of the operation of modeling the temporal dependency and the result of the operation of extracting correlations between the plurality of features.

[0005] A system according to one embodiment of the present disclosure comprises one or more processors; and one or more memories that collectively store instructions that cause the system to perform operations when executed by the one or more processors, wherein the operations may include: an operation of acquiring time series data; an operation of generating a plurality of randomly extracted data by performing a plurality of random extractions on the time series data; an operation of performing a Temporal Attention operation that models the temporal dependency of the plurality of randomly extracted data; an operation of performing an Inter-Feature Attention operation that extracts correlations between a plurality of features of the plurality of randomly extracted data; and an operation of generating at least one prediction data based on the result of the Temporal Attention operation and the result of the Inter-Feature Attention operation.

[0006] In one embodiment, the time series data includes D unit data, and the operation of generating the plurality of randomly extracted data includes: the operation of generating D unit data by dividing the time series data based on a time axis; and the operation of generating the plurality of randomly extracted data by performing the operation of randomly selecting S unit data from the D unit data multiple times, wherein S may be a value greater than 1 and less than D.

[0007] In one embodiment, the operation of performing the time-axis attention operation and the operation of performing the feature-to-feature attention operation are performed sequentially, and the operation of performing the feature-to-feature attention operation may include the operation of extracting the correlation between the plurality of features with respect to the result of the time-axis attention operation.

[0008] In one embodiment, the operation of generating at least one prediction data includes the operation of generating at least one prediction data by sequentially performing the operation of performing an operation of modeling the temporal dependency and the operation of performing an operation of extracting the correlation between the plurality of features, and the operation of performing the operation of extracting the correlation between the plurality of features may include the operation of extracting the correlation between the plurality of features with respect to the result of the operation of modeling the temporal dependency.

[0009] In one embodiment, the operation of performing the attention operation between features may include: an operation of transforming the array axis of the result of the time axis attention operation; and an operation of performing the attention operation between features based on the feature axis for the data in which the array axis has been transformed.

[0010] In claim 1, the operation of performing an operation to extract correlations between the plurality of features may include: an operation to transform the array axis of the result of the operation to model the temporal dependency; and an operation to perform an operation to extract correlations between the plurality of features based on the feature axis for the data in which the array axis has been transformed.

[0011] In one embodiment, the operations further include: an operation of generating a plurality of tokens by dividing the plurality of randomly extracted data based on a time axis; and an operation of adding a position embedding representing temporal position information and characteristic position information to each of the plurality of tokens, wherein the time axis attention operation and the feature attention operation may be performed based on the plurality of tokens to which the position embedding has been added.

[0012] In one embodiment, the operation for modeling the temporal dependency and the operation for extracting the correlation between the plurality of features may be performed based on a plurality of tokens to which the position embedding is added.

[0013] In one embodiment, the operation of generating the at least one prediction data comprises: an operation of calculating an individual prediction value for each of the plurality of randomly extracted data; and calculating the average of the individual prediction values ​​for each of the plurality of randomly extracted data to generate the at least one prediction data.

[0014] In one embodiment, the operation of generating at least one prediction data may include the operation of generating at least one prediction data by inputting the result of the time-axis attention operation and the result of the feature-to-feature attention operation into a multi-layer perceptron (MLP) module.

[0015] In one embodiment, the operation of generating at least one prediction data may include the operation of generating at least one prediction data by inputting the result of an operation modeling the temporal dependency and the result of an operation extracting the correlation between the plurality of features into a Multi-Layer Perceptron (MLP) module.

[0016] A method for generating prediction data performed by at least one processor according to one embodiment of the present disclosure may include: acquiring time series data; generating a plurality of randomly extracted data by performing a plurality of random extractions on the time series data; performing a temporal attention operation to model the temporal dependency of the plurality of randomly extracted data; performing an inter-feature attention operation to extract correlations between a plurality of features of the plurality of randomly extracted data; and generating at least one prediction data based on the result of the temporal attention operation and the result of the inter-feature attention operation.

[0017] In one embodiment, the time series data includes D unit data, and the step of generating the plurality of random extraction data includes: a step of generating D unit data by dividing the time series data based on a time axis; and a step of generating the plurality of random extraction data by performing a multiple operation of randomly selecting S unit data from among the D unit data, wherein S may be a value greater than 1 and less than D.

[0018] In one embodiment, the step of performing the time-axis attention operation and the step of performing the feature-to-feature attention operation are performed sequentially, and the step of performing the feature-to-feature attention operation may include the step of extracting the correlation between the plurality of features with respect to the result of the time-axis attention operation.

[0019] In one embodiment, the step of performing the attention operation between features may include: a step of transforming the array axis of the result of the time axis attention operation; and a step of performing the attention operation between features based on the feature axis for the data in which the array axis has been transformed.

[0020] In one embodiment, the method further comprises the step of generating a plurality of tokens by dividing the plurality of randomly extracted data based on a time axis; and the step of adding a position embedding representing temporal position information and characteristic position information to each of the plurality of tokens, wherein the time axis attention operation and the feature attention operation may be performed based on the plurality of tokens to which the position embedding has been added.

[0021] In one embodiment, the step of generating at least one prediction data may include: a step of calculating an individual prediction value for each of the plurality of randomly extracted data; and a step of generating at least one prediction data by calculating the average of the individual prediction values ​​for each of the plurality of randomly extracted data.

[0022] In one embodiment, the step of generating at least one prediction data may include the step of generating at least one prediction data by inputting the result of the time-axis attention operation and the result of the feature-to-feature attention operation into a multi-layer perceptron (MLP) module.

[0023] One embodiment of the present disclosure includes a program stored on a recording medium to execute a method according to one embodiment of the present disclosure on a computer.

[0024] One embodiment of the present disclosure includes a computer-readable recording medium having a program for executing a method according to one embodiment of the present disclosure on a computer.

[0025] One embodiment of the present disclosure includes a computer-readable recording medium that records a database used in one embodiment of the present disclosure.

[0026] According to one embodiment of the present disclosure, efficient and highly accurate prediction data can be determined using multivariate time series input data.

[0027] FIG. 1 illustrates an example of a block diagram of a computing system performing a time series forecasting method according to one embodiment of the present disclosure.

[0028] FIG. 2 illustrates an example of a block diagram of a computing device, which is one of the components of a computing system performing a time series forecasting method according to one embodiment of the present disclosure.

[0029] FIG. 3 illustrates an example of a block diagram in another aspect of a computing device, which is one of the components of a computing system performing a time series forecasting method according to one embodiment of the present disclosure.

[0030] FIG. 4a is a diagram illustrating a time series forecasting method of a transformer model in which one observation becomes one token according to one embodiment of the present disclosure.

[0031] FIG. 4b is a diagram illustrating a time series prediction method for a segment-based transformer model according to one embodiment of the present disclosure.

[0032] FIG. 5 is a diagram showing a time series prediction model according to one embodiment of the present disclosure.

[0033] FIG. 6 is a graph showing the correlation between the number of repetitions of random division and the error according to one embodiment of the present disclosure.

[0034] FIG. 7 is a diagram illustrating a method for generating random extraction data according to one embodiment of the present disclosure.

[0035] FIG. 8 is a block diagram of a time series forecasting system according to one embodiment of the present disclosure.

[0036] To clarify the technical concept of the present disclosure, embodiments of the present disclosure will be described in detail with reference to the attached drawings. In describing the present disclosure, detailed descriptions of related known functions or components will be omitted if it is determined that such detailed descriptions would unnecessarily obscure the essence of the present disclosure. Components having substantially the same functional configuration among the drawings have been assigned the same reference numerals and symbols as much as possible, even if they are shown in different drawings. For convenience of explanation, devices and methods will be described together where necessary. Each operation of the present disclosure does not necessarily have to be performed in the order described and may be performed in parallel, selectively, or individually.

[0037] The terms used in the embodiments of this disclosure have been selected to be as widely used and general as possible, taking into account the functions of this disclosure; however, these terms may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been selected at the applicant's discretion, and in such cases, their meanings will be described in detail in the description of the relevant embodiments. Therefore, terms used in this specification should be defined not merely by their names, but based on their meanings and the overall content of this disclosure.

[0038] Throughout this disclosure, singular expressions may include plural expressions unless the context clearly indicates otherwise. Terms such as “comprising” or “having” are intended to specify the presence of features, numbers, steps, actions, components, parts, or combinations thereof, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof. That is, throughout this disclosure, when a part is described as “comprising” a certain component, it means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0039] Expressions such as "at least one" modify the entire list of components and do not modify the components of the list individually. For example, "at least one of A, B, and C" and "at least one of A, B, or C" refer to only A, only B, only C, both A and B, both B and C, both A and C, all of A, B, and C, or any combination thereof.

[0040] Additionally, terms such as “...part,” “...module,” etc., as described in this disclosure refer to a unit that processes at least one function or operation, and may be implemented in hardware or software, or a combination of hardware and software.

[0041] Throughout the entire disclosure, when a part is described as being “connected” to another part, this includes not only cases where they are “directly connected” but also cases where they are “electrically connected” with other elements interposed between them. Furthermore, when a part is described as “comprising” a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0042] As used throughout this disclosure, the expression “configured to” may be replaced, depending on the context, with, for example, “suitable for,” “having the capacity to,” “designed to,” “adapted to,” “made to,” or “capable of.” The term “configured to” may not necessarily mean only “specifically designed to” in hardware. Instead, in some situations, the expression “system configured to” may mean that the system is “capable of” together with other devices or components. For example, the phrase “a processor configured (or set) to perform A, B, and C” may mean a dedicated processor for performing said operations (e.g., an embedded processor), or a generic-purpose processor (e.g., a CPU or an application processor) capable of performing said operations by executing one or more software programs stored in memory.

[0043] Throughout the entire disclosure, the expression [N : M] is a set of integers from N to M, wherein N is included and M is not included. That is, [N : M] may mean {N, N+1, ..., M-1}.

[0044] FIG. 1 illustrates an example of a block diagram of a computing system performing a time series forecasting method according to one embodiment of the present disclosure.

[0045] Referring to FIG. 1, a computing system (1000) for performing target prediction according to one embodiment of the present disclosure includes a user computing device (110), a training computing system (150), and a server computing system (130), and each device and system is connected to communicate through a network (170).

[0046] According to one embodiment, 1) a user computing device (110) can perform a target prediction method using a local or / and external machine learning model (120) or a machine learning model (140) provided by a server.

[0047] Additionally, according to another embodiment, 2) a server computing system (130) communicating with a user computing device (110) may provide a target prediction service to the user computing device (110) on an application or / and the web in response to a request from a user through the user computing device (110).

[0048] In addition, according to another embodiment, 3) a user computing device (110) and a server computing system (130) may perform at least a part of the method of performing target prediction in conjunction with each other to provide a target prediction service to the user.

[0049] Additionally, according to various embodiments of the present disclosure, a user computing device (110) and / or a server computing system (130) may learn a machine learning model (120 / 140) that is performed on target prediction through interaction with a training computing system (150) that is communicatedly connected via a network (170). In this case, the training computing system (150) may be separate from the server computing system (130) or may be part of the server computing system (130).

[0050] In some embodiments, the training computing system (150) may be part of the server computing system (130) or part of the user computing device (110).

[0051] In the following description, the method is described based on connecting to a server computing system (130) via a user computing device (110) to execute a target prediction task, collecting and analyzing data necessary for target prediction directly from the server computing system (130) or using a language model from a separate server, and performing a target forecast prediction based on the collected and analyzed data. However, it can be understood that if a part of the process described as being performed in the server computing system (130) is performed in the user computing device (110), it is naturally included in the description of the present invention.

[0052] The user computing device (110) may include all other types of computing devices, such as a smartphone, a mobile phone, a digital broadcasting device, a PDA (personal digital assistants), a PMP (portable multimedia player), a desktop, a wearable device, an embedded computing device and / or a tablet PC.

[0053] The user computing device (110) includes at least one processor (111) and memory (112). Here, the processor (111) may be composed of at least one of a central processing unit (CPU), a graphics processing unit (GPU), ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), controllers, microcontrollers, microprocessors, and / or electrical units for performing other functions, or a plurality of electrically connected processors.

[0054] The memory (112) may include one or more non-transient / transient computer-readable storage media such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, and combinations thereof, and may include web storage of a server that performs memory storage functions on the internet. This memory (112) may store data and instructions necessary for the at least one processor (111) to perform the operation of an application for performing target prediction.

[0055] In one embodiment, the user computing device (110) may store at least one machine learning model (120). For example, the user computing device (110) may be various machine learning models, such as multiple neural networks (e.g., deep neural networks) that perform predictions on a target based on structured / quantitative data, or other types of machine learning models including non-linear models and / or linear models, and may be composed of a combination thereof.

[0056] For example, the prediction model may include linear regression, decision tree, random forest, gradient boosting pre-trained language model or / and deep learning model. The neural network model may include feedforward neural networks, recurrent neural networks (e.g., LSTM), convolutional neural networks, Transformers or / and combinations thereof.

[0057] In particular, in one embodiment according to the present disclosure, a partial feature-based transformer model (PMformer) that selects only some features from multivariate time series data and uses them as input may be stored or executed, and this model can simultaneously ensure the efficiency and accuracy of prediction by learning only partial relationships between the selected features.

[0058] In one embodiment, the user computing device (110) can perform time series prediction in real time by utilizing a locally stored time series prediction model. In another embodiment, the user computing device (110) transmits a time series prediction request to a server computing system (130), and the server processes time series data based on at least one prediction model and provides the prediction result to the user device.

[0059] In one embodiment, the user computing device (110) may include one or more input components (121) for detecting user input. For example, the input components (121) may include a touch sensor, an image sensor, a microphone, a button, a mouse, a keyboard, etc., and may include an interface capable of receiving input from an external controller (e.g., a mouse, a keyboard, etc.).

[0060] A server computing system (130) may include at least one processor (131) and memory (132). Here, the processor (131) may be composed of one or more electrical computing devices for performing time series forecasting operations, such as a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processing unit (DSP), or an FPGA, a microprocessor, etc. The memory (132) may include a non-transient or transient storage medium such as RAM, ROM, flash memory, or a magnetic disk, and may store instructions or data for the execution and inference of a time series forecasting model, in particular a partial feature-based Transformer model (PMformer), executed by the processor (131).

[0061] The prediction model (140) may include at least one of a feedforward neural network, a recurrent neural network (LSTM), a convolutional neural network (CNN), or a transformer-based model, and in one embodiment of the present disclosure, a shared transformer model (PMformer) that performs partial time series prediction based on a randomly selected set of features may be included.

[0062] In one embodiment, the server computing system (130) may be composed of a single device or a plurality of computing devices, and may be configured to operate according to a sequential or parallel computing architecture. Additionally, a distributed processing system may be configured with a plurality of networked devices.

[0063] The server computing system (130) may further include a data store (data store). This data store may be composed of a relational database, a NoSQL database, a data warehouse, or a local file system, and may be used to store and manage time-series input data, prediction results, training logs, etc.

[0064] The training computing system (150) may include at least one processor (151) and memory (152). This system may be used as an environment for performing pre-training or retraining of a PMformer model. The memory (152) may store a dataset, loss function, optimization algorithm, parameters, etc., for model training.

[0065] For example, the training computing system (150) can update the parameters of the PMformer by executing a learning algorithm including backpropagation. This process may include iterative learning to improve time series forecasting performance, and generalization techniques such as dropout and weight reduction may be applied.

[0066] The model trainer (160) is a logical component that executes this learning and can be implemented in the form of hardware, firmware, or software. For example, the model trainer can be executed by a processor after loading a learning program file stored in a storage device into memory.

[0067] The system can be connected via a wired / wireless network (170) for communication between a user computing device (110) and a server computing system (130). This network may include various forms such as the Internet, LAN, wireless LAN, WAN, and mobile communication network, and data transmission may be performed based on standard communication protocols such as TCP / IP, HTTP, and SSL.

[0068] Additionally, the user computing device (110) may store a model to be used in each process and a prompt template that serves as the basis for input to the model in order to perform at least part of the process for target prediction through a large-scale language model (LLM).

[0069] For example, the user computing device (110) may store: 1) a prompt for generating a query from user input, 2) a prompt for determining the relationship between the target and the target influence variable, 3) a prompt for identifying raw data associated with the determined relationship, 4) a prompt template for quantifying unstructured data, etc.

[0070] That is, in one embodiment, the user computing device (110) can perform a target prediction based on the received data by requesting the execution of some execution steps in the target prediction task to a language model of an external server via a prompt, etc.

[0071] In another embodiment, for a target prediction task requested through a user computing device (110), the server computing system (130) may perform a target prediction using at least one machine learning model (140) and a machine learning model of another server and provide the predicted data to the user computing device (110).

[0072] FIG. 2 illustrates an example of a block diagram of a computing device, which is one of the components of a computing system performing a time series forecasting method according to one embodiment of the present disclosure.

[0073] Referring to FIG. 2, a computing device (100) included in a user computing device (110), a server computing system (130), and a training computing system (150) includes a plurality of applications (e.g., applications 1 to N). Each application may include a machine learning library. For example, the applications may include a futurecasting application, a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, a chat-bot application, a separate futurecasting application, etc. The computing device (100) may include a plurality of applications (e.g., applications 1 to N), and each application may include a machine learning library and a model execution environment for performing machine learning-based time series processing functions. For example, the computing device (100) may include a futurecasting application, that is, a specialized application for time series forecasting, and the application may perform time series forecasting based on the PMformer model of the present invention.

[0074] In one embodiment, the computing device (100) may include a model trainer (160) for pre-training or retraining a PMformer model. The model trainer (160) may learn or update parameters based on input time series data and save the learned model to be used for real-time or non-real-time prediction tasks.

[0075] Each application included in the computing device (100) can communicate with sensors, device state managers, or external input devices within the device via an Application Programming Interface (API). In one embodiment, each application can interface with device components, such as receiving sensor data or state data or transmitting prediction results to an output device, via a public or private API.

[0076] FIG. 3 illustrates an example of a block diagram in another aspect of a computing device, which is one of the components of a computing system (1000) that performs a time series prediction method according to one embodiment of the present disclosure.

[0077] Referring to FIG. 3, the computing device (200) includes a plurality of applications (e.g., Application 1 to Application N). Each application can communicate with a central intelligence layer. For example, the applications may include a time series forecasting application, an energy management application, a logging and analysis application, an image processing application, a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. In one embodiment, each application can communicate with the central intelligence layer (and the model stored therein) using an API (e.g., a common API across all applications).

[0078] The central intelligence layer includes one or more machine learning models for time series forecasting (e.g., PMformer models) and may share them among multiple applications or provide them independently to each. For example, as illustrated in FIG. 3, each machine learning model and at least some thereof may be provided to each application and managed by the central intelligence layer. In one embodiment, the central intelligence layer may be integrated as part of the operating system or implemented as a separate logical layer and may perform the role of passing input time series data to the corresponding model to return a forecast result. Two or more applications may share a single machine learning model. For example, in some embodiments, the central intelligence layer may provide a single model to all applications. In some embodiments, the central intelligence layer may be included within the operating system of the computing device (200) or implemented otherwise.

[0079] Additionally, the central intelligence layer can communicate with the central device data layer. The central device data layer may be a centralized data store for the computing device (200). As illustrated in FIG. 3, the central device data layer may communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, and / or additional components. In some embodiments, the central device data layer may communicate with each device component using an API (e.g., a private API). For example, the central device data layer may store sensor data, device state information, external environment information, etc., stored within the computing device, and provide this as input data required for time series forecasting. Each device component (e.g., a sensor, a state manager, etc.) may communicate with the corresponding data layer through a private API, etc.

[0080] The technology described herein may refer to servers, databases, software applications, and other computer-based systems, as well as actions taken and information transmitted to or from said systems. It will be recognized that the inherent flexibility of computer-based systems allows for a wide range of possible configurations, combinations, division of tasks, and functionality between and from components. For example, the processes described herein may be implemented using a single device or component or multiple devices or components operating in combination. Databases and applications may be implemented in a single system or in a distributed system across multiple systems. Distributed components may operate sequentially or in parallel.

[0081] Time series forecasting can involve fundamental machine learning tasks aimed at predicting future events based on past observations. These forecasting problems often require long-term forecasting and can utilize various variables. For example, stock price forecasting may require predicting multiple market values ​​over a long time axis. In these multivariate long-term time-series forecasting (M-LTSF) problems, it is important to capture both the long-term temporal dependencies between past and future events and the dependencies between the characteristics of different variables.

[0082] In recent years, many deep neural architectures, including linear models, state-space models, and Recurrent Neural Networks (RNNs), have been developed for the M-LTSF problem. Among them, the Transformer model is a neural network that learns context and semantics by tracking relationships within sequential data, such as words in a sentence. Transformer models have demonstrated remarkable performance in various fields, such as language and image processing, and due to their ability to capture long-term relationships, they are also being studied in the field of multivariate long-term time series forecasting. For example, as shown in Fig. 4a, a Transformer model in which a single observation becomes a single token has been used in the field of time series forecasting. In recent research, a segment-based Transformer model has been proposed in which each token is represented as a continuous group of observations rather than a single observation, as shown in Fig. 4b. However, in the case of self-attention in segment-based transformer models, one segment is converted into one token; while prediction performance improves as the segment becomes more subdivided, the number of tokens increases significantly when subdivided, thereby greatly increasing the computational cost of attention. Additionally, as shown in FIG. 4b, in inter-feature attention, which seeks associations between features, prediction can be performed significantly inefficiently when the number of features is very large. To address these problems, one embodiment of the present disclosure aims to provide a time series prediction method that maintains performance while being less subdivided, and maintains performance even in inter-feature attention with a large number of features. The transformer model provided by one embodiment of the present disclosure may be referred to as PMformer (Partial-Multivariate Transformer) throughout the present disclosure.

[0083] FIG. 5 is a diagram showing a time series prediction model according to one embodiment of the present disclosure.

[0084] Referring to FIG. 5, the processor can acquire time-series data (510). The time-series data (510) may include multivariate time-series data that records the changing patterns of multiple variables or features over time. For example, if the time-series data (510) has a matrix structure of size T / X / D, the time-series data (510) is x, which is the observed value of the d-th feature at the t-th time point. t,d It can be defined as a data set including . Here, D represents the total number of features, and T represents the time series length (Historical look-back window) of the observed data from the past to the present.

[0085] In addition, the time series data (510) acquired by the processor is an observation x during a past period ([0, T]). [0:T],d It includes x, which is the predicted data for the future period ([T, T+ τ]) to be predicted. [T:T+τ],d It can be used as an input value to generate. Here, τ represents the length of the future time point to be predicted.

[0086] In one embodiment, the feature (D) may include various indicators representing the state of the target system. For example, the feature (D) may include power indicators including power supply amount and load information, weather indicators including temperature and humidity, or traffic indicators representing road occupancy. Alternatively, in relation to the naphtha cracking process, the inventory amount of each raw material storage tank, property information, production volume by product, and raw material price history may be included as the feature (D).

[0087] In one embodiment, the processor may generate multiple randomly extracted data (520) by performing multiple random extractions on time series data (510). This random extraction process may be performed to build a partial-multivariate learning and prediction environment that captures only some of the relationships among the complex relationships between all features.

[0088] Specifically, the processor may randomly select a subset of features having a preset size S (1 < S < D) from multivariate time series data (510) in which the total number of features is D. In order to resolve noise or overfitting problems that may occur when an existing complete-multivariate model attempts to calculate the correlation between all features (D) at once, one embodiment of the present disclosure may intentionally extract only S of the features to form a sub-scenario.

[0089] In one embodiment, the total number D is not necessarily limited to representing only the number of features of the data, but may refer to the sum of unit data that constitute the constituent units of the data to be processed by the system. Time series data may include multiple D unit data, and the processor can ensure computational efficiency and model robustness by extracting S of these to form a subset. For example, the processor can generate D unit data by dividing the acquired time series data according to a pre-set interval or number based on the time axis. This enables fine sampling not only on the feature axis of multivariate data but also on the time axis, thereby providing a general-purpose learning environment that is not limited to specific time intervals or specific feature combinations of the data.

[0090] The processor can generate multiple randomly extracted data by performing the operation of randomly selecting S unit data from D generated unit data multiple times. In this case, the number of extracted unit data S is set to a value greater than 1 and less than D (1 < S < D), thereby enabling partial data modeling that utilizes only a portion of the total data. This configuration prevents computational complexity and overfitting issues that may arise when the model attempts to calculate all correlations present in the entire data at once, and allows for more precise estimation of potential patterns in the entire data by iteratively learning from S data of various randomly selected combinations. Furthermore, even if some unit data is lost or contains abnormal noise, the prediction results generated from subsets of other combinations are averaged to compensate, thereby significantly improving the overall prediction stability of the system.

[0091] The processor can train a flexible model that does not rely solely on specific feature combinations by generating multiple randomly extracted data (520) by randomly sampling combinations of different features during each training iteration. This can induce the model to extensively learn partial dependencies between various features.

[0092] In the inference stage, instead of using only a single combination of features, the processor may independently generate multiple randomly extracted data (520) containing different feature subsets through multiple random extractions. Subsequently, the processor may perform individual predictions for each extracted data and synthesize (e.g., average) the results, thereby offsetting errors that may occur in individual predictions and bringing about an ensemble effect that increases the accuracy of the final output value.

[0093] In a random extraction process according to one embodiment, the number of features S extracted may be a fixed value, but may also be set as a certain ratio (e.g., 50%) to the total number of features D. The processor may convert each of the multiple random extraction data (520) obtained through this extraction process into an independent input token sequence and provide it as an input for a subsequent operation. In one embodiment, the operation may include an operation for modeling temporal dependency, an operation for extracting correlations between multiple features, etc. For example, the operation for modeling temporal dependency may include a time-axis attention operation, and the operation for extracting correlations between multiple features may include attention between features.

[0094] In this disclosure, partial multivariate sampling is described using random sampling as an example, but it is not limited to random sampling. Partial multivariate sampling may include feature-based sampling, time-interval-based sampling, feature and time-interval-based sampling, sampling with and without duplicates, sampling without duplicates, priority sampling of important features, sampling excluding missing features, adaptive sampling during inference, etc. Additionally, the number of samples can be dynamically adjusted.

[0095] For example, feature-based sampling randomly selects a subset of S features from a total of D features to intensively learn partial correlations between variables, while time-interval-based sampling selects S data points from D unit data points partitioned along the time axis to reduce computational complexity while enabling the even learning of overall time-series patterns. Furthermore, feature and time-interval-based sampling can construct a partial multivariate environment that integrates spatiotemporal context by extracting a 2D submatrix composed of data from specific time intervals for specific features. These dimension-based sampling strategies can maximize computational efficiency by cleverly limiting the scope of learning required by the model while maintaining the inherent structure of the data.

[0096] As another example, duplicate-allowing sampling can be advantageous for increasing the learning intensity of specific key patterns by ensuring that the same data unit is included multiple times within a subset. Additionally, non-duplicate sampling, when combined with a random partitioning algorithm, can guarantee that all data participates in model training at least once without omission within a single training iteration.

[0097] As another example, Key Feature First Sampling and Missing Feature Exclusion Sampling can be utilized. Key Feature First Sampling is a method that directly links prediction accuracy to the results by assigning high selection probability weights to key indicators that contribute significantly to the forecast outcome, such as power consumption data in power demand forecasting. Missing Feature Exclusion Sampling, on the other hand, maintains the predictive stability of the model even in environments with incomplete real-world data by immediately excluding features where observations are missing—such as sensor failures—from the sampling candidate pool. Such flexible sampling can serve as a highly practical solution in real-world industrial settings with significant data noise, such as finance, healthcare, and transportation.

[0098] In one embodiment, the processor may perform a segmentation process on a plurality of randomly extracted data (520) extracted using a segmentation module (530). Specifically, the processor may perform a segmentation process on a preset number (N) of past observation data (T) for each feature. SThe data can be divided into intervals and each interval can be encoded into latent tokens through a linear layer. At this time, the method of generating tokens is not limited to one-dimensional time axis partitioning. According to an embodiment, the processor may perform tokenization by partitioning the data into a 2D patch form by simultaneously considering the time axis and the feature axis. In this process, absolute or relative position embeddings indicating temporal position information (order on the time axis) and feature position information (corresponding to which variable) are added to each token, so that the learning model can simultaneously recognize the order of the time series and the relationship between features.

[0099] In one embodiment, the processor may perform a Temporal Attention operation that models the temporal dependency of a plurality of randomly extracted data by a Temporal Attention module (540). The Temporal Attention module (540) can capture the pattern of change of data over time within each feature constituting the plurality of randomly extracted data (520). Here, Temporal Attention is described as an example, but the Temporal Attention module (540) may be replaced with another model that models temporal dependency.

[0100] In one embodiment, the time-axis attention module (540) may receive feature-specific tokens generated by the segmentation module (530) as input. The time-axis attention module (540) may calculate the correlation between multiple segments arranged along the time axis independently for each feature. For example, the time-axis attention module (540) can effectively model long-term trends or periodic seasonal patterns existing within the time-series data by calculating an attention score indicating how segment data at a specific point in time has a relationship with segment data at other points in time in the past or future.

[0101] In one embodiment, the time-axis attention module (540) may operate as part of a shared transformer structure. Instead of using a separate model for each of the multiple randomly extracted data (520), the processor can increase learning efficiency and achieve lightweighting of the model by repeatedly applying the time-axis attention module (540) that shares the same parameters.

[0102] This temporal attention operation is performed only on features within a randomly extracted subset, rather than considering all complete multivariate features. This offers the advantage of precisely learning the unique temporal characteristics of each feature while minimizing the influence of unnecessary noise. Through this, the processor can generate a high-dimensional feature vector reflecting the temporal contextual information of each token and provide it as input for subsequent inter-feature attention operations.

[0103] In one embodiment, the processor may perform an inter-feature attention operation to extract correlations between multiple features of multiple randomly extracted data by an inter-feature attention module (550). The inter-feature attention module (550) can model the interdependence of multivariate time series data by capturing how different features influence each other within the same time interval. Here, inter-feature attention is described as an example, but the inter-feature attention module (540) can be replaced with another model that extracts correlations between multiple features.

[0104] In one embodiment, the feature attention module (550) may receive tokens reflecting temporal context information as input through the previous step time-axis attention operation. The processor may perform multi-head self-attention between features within a randomly sampled subset while keeping each time segment position fixed. Through this, the system can calculate and represent the dynamic relationship between variables, such as the effect of a change in feature A on feature B at a specific point in time, using learned weights.

[0105] According to one embodiment, the attention operation between features can be performed only on a feature subset having a preset size S (1 < S < D) instead of processing the entire feature set (D) at once. This reduces the complexity of the existing method of calculating relationships between all features to O(D 2 By significantly reducing the O(SD) level, efficient inference can be achieved even in time-series data containing large-scale variables. Furthermore, the feature attention operation can improve the model's generalization performance by excluding potential noise that may occur between irrelevant features and focusing on learning only significant correlations within randomly sampled subsets. Additionally, even if some feature data is missing, the feature attention operation can continue prediction using subsets that exclude the missing feature during the random sampling process, thereby maintaining stable performance even in real-world environments with incomplete data.

[0106] In one embodiment, the processor generates a final representation in which both temporal information and inter-feature correlation information are integrated by undergoing temporal attention and inter-feature attention, and can generate at least one prediction data (560) using this. Throughout the disclosure, the operation for modeling temporal dependency may include a Temporal Attention operation, and the operation for extracting correlations between a plurality of features may include an Inter-Feature Attention operation.

[0107] That is, the processor can generate at least one prediction data (560) based on a final representation in which temporal and characteristic correlation information generated through the preceding operations is integrated, and determine this as the final output value.

[0108] First, in the prediction data generation stage, the processor can derive future observation data for each subset. Specifically, the processor's decoder receives feature-specific latent tokens generated in the last transformer block and passes them through a single linear layer to derive future observation data for each feature. At this time, since each subset data reflects only the correlation between some randomly sampled features (S), the processor can generate independent prediction results for each individual subset.

[0109] In the ensemble and final result determination stages, the processor can integrate the generated multiple prediction results to produce a final output value. An inference technique according to one embodiment of the present disclosure performs an inference process N based on random partitioning. I Perform the process repeatedly, and the N obtained therefrom IThe final result can be obtained by performing an average operation on the output values. However, the combination method for obtaining the final result is not limited to a simple arithmetic mean; the processor may calculate a weighted average by assigning different weights based on the data reliability of each subset or the importance of the included features, or extract a median value to exclude the influence of extreme outliers. Additionally, the processor may further enhance the robustness of the ensemble by applying a voting method that selects the optimal value through majority modeling among multiple prediction results.

[0110] According to this ensemble (averaging) operation, the probability that mutually significant features are included within a specific subset is the number of trials, N. I As increases, it increases together. Therefore, by having the processor sample multiple subsets to repeat predictions and average the results, the likelihood that highly correlated feature combinations contribute to the prediction can be maximized, which ultimately ensures superior prediction accuracy and robustness compared to a single prediction model.

[0111] In one embodiment, the number of repetitions N I can be adjusted according to user settings, and according to the experimental results of Fig. 6, N I As ε increases, the test error (MSE) tends to decrease monotonically; therefore, the system considers the balance between resource efficiency and accuracy to determine the optimal number of iterations N. I can decide.

[0112] In one embodiment, the processor may perform additional data processing by a Multi-Layer Perceptron (MLP) module after at least one of a time-axis attention operation and a feature-to-feature attention operation. The MLP module may perform the role of non-linearly transforming and refining temporal and feature correlation information extracted through the attention module. The MLP module may receive an intermediate representation in which the attention operation is completed as input and generate a final output value.

[0113] In one embodiment, the MLP module can perform operations on both the feature axis and the time axis. Through this, the MLP module can reflect the multidimensional characteristics of time-series data. By learning complex non-linear patterns of the data through linear layers and activation functions within the MLP, the processor can complement the expressive power of the model, which may be insufficient with attention mechanisms alone. Furthermore, the MLP module according to one embodiment can maximize the learning efficiency of the model by sharing and using the same parameters for all subsets within a shared transformer structure.

[0114] Through this configuration, the processor can more precisely refine potential tokens reflecting temporal dependencies and correlations between features, and ultimately derive more accurate future observations in the prediction data generation unit.

[0115] FIG. 7 is a diagram illustrating a method for generating random extraction data according to one embodiment of the present disclosure.

[0116] Referring to FIG. 7, the processor can generate multiple randomly extracted data (650, 660, 670). This process can serve as the basis for partial multivariate modeling to efficiently learn partial dependencies among all features. The processor can identify an entire time-series data set containing multiple features (610, 620, 630, 640, 650). At this time, when the total number of features is defined as D, the processor can determine the size S of a subset containing only some of the total features (1 < S < D). In the example of FIG. 7, the total number of features D is 5, and the size S of the sampled subset can be set to 3.

[0117] The specific method by which the processor generates randomly extracted data (650, 660, 670) can be performed using random sampling, random partitioning, etc.

[0118] First, the random sampling method can refer to a method in which the processor constructs a subset F by completely randomly selecting S features from the entire feature set ([0:D]) in each iteration. While the random sampling method can generate various feature combinations even without prior knowledge, cases may occur where some features are selected redundantly or are not selected at all in a specific iteration.

[0119] Next, in the random partitioning method, the processor divides the entire feature D into N non-overlapping features. U (N UThis can mean a method of dividing into (D / S) subsets. For example, when extracting feature combinations from the entire set of features as shown in Fig. 7, data omission can be prevented and learning efficiency maximized by ensuring that each feature is input into the prediction model at least once. If D is not divisible by S, the processor can configure all subsets to have the same size S by randomly duplicating some features to supplement them.

[0120] The first random extraction data (650), the second random extraction data (660), and the third random extraction data (670) generated through this process can each be treated as independent input instances. By inputting the multiple random extraction data (650, 660, 670) generated in this way into a shared neural network sequentially or in parallel, the processor can learn partial correlations between features and predict future values.

[0121] In the inference step according to one embodiment, the processor performs this random extraction process N I By averaging multiple prediction results obtained through repeated iterations, a final output value that is robust to missing values ​​and has improved prediction accuracy can be derived.

[0122] After generating randomly extracted data according to FIG. 7, the processor can simultaneously learn temporal dependencies and correlations between features using a shared transformer structure. According to one embodiment, a single neural network may be shared by all sampled subsets. Instead of building separate prediction models for each of the first to third randomly extracted data (650, 660, 670) generated in the example of FIG. 7, a single model having the same set of parameters may be shared and used. This allows the model to repeatedly experience different combinations of features, thereby inducing it to learn universal dynamic relationships inherent in the entire data without overfitting to a specific set of features.

[0123] In one embodiment, the time-axis attention module can independently capture temporal dependencies for the features included in each subset. The time-axis attention module can perform Multi-Head Self-Attention (MHSA) on the token sequences for each feature within the sampled subset. The attention weights (Query, Key, Value parameters) used can be applied equally to all data regardless of the type of subset or the index of the feature. Through this, the model can possess a common ability to extract the unique time-series patterns (e.g., periodicity, trend, etc.) of a variable, regardless of which variable is input.

[0124] In one embodiment, the inter-feature attention module can model the correlation between S features extracted within the same time interval. The inter-feature attention module can calculate the interaction between tokens arranged along the feature axis at each time segment position. Although the combination of input features, such as 650, 660, and 670 in FIG. 7, varies each time, the parameter analyzing the correlation between them can be fixed as one and shared. Consequently, the model can learn a generalized rule on 'how to calculate partial dependencies between them when arbitrary variables are given,' rather than individually memorizing the 'relationship between feature A and B.'

[0125] A processor according to one embodiment can model complex dependencies of time-series data step by step by sequentially performing time-axis attention operations and feature-to-feature attention operations on a plurality of randomly extracted data. For example, the processor can first perform time-axis attention operations within each PMformer block to capture temporal dependencies within individual features. The feature-to-feature attention module can receive as input result data in which temporal context information is reflected after the time-axis attention operations are completed. However, the order of operations of the present disclosure is not limited thereto. According to one embodiment, the processor may adopt an architecture in reverse order in which feature-to-feature attention operations are performed first, followed by time-axis attention operations, depending on the characteristics of the data or system load. Furthermore, the processor may generate final prediction data by simultaneously performing operations to model temporal dependencies and operations to extract correlations between features in parallel on independent branches, and then combining the results of the two operations (e.g., concatenation, element-wise addition, or merging after weighting). These changes in order and parallel processing are intended to maximize the model's flexibility in accordance with the domain characteristics of the time-series data. By extracting correlations between features after temporal information has been integrated in this manner, the model can move beyond mere similarity between time points and more precisely identify dependencies between variables based on the dynamic flow of each variable.

[0126] Furthermore, the processor can perform array axis transformation operations for efficient data processing during the process of performing attention operations between features. For tensor data resulting from time-axis attention operations, the processor can align the time dimension and feature dimension axes by swapping or rearranging them so that the feature axis becomes the primary target of the operation. For the data with transformed array axes, the processor can perform Multi-Head Self-Attention (MHSA) based on the feature axis while keeping each time point fixed. Through this axis transformation process, the processor can freely switch the direction of operations even within the same shared transformer structure, which can provide a structural foundation that enables the model to learn integrated row and column information of multivariate time series without omission.

[0127] A partial multivariate time series forecasting system according to one embodiment of the present invention can be applied to forecasting problems in various industrial fields involving multiple interrelated variables. By performing learning and inference by sampling a subset containing only a portion of the total features, the system can simultaneously ensure forecasting accuracy and robustness against missing data while reducing computational resources.

[0128] The fields in which an embodiment of the present disclosure can be utilized are as shown in [Table 1] below.

[0129] Application Fields Technical Objective Examples of Input Data Examples of Output Data Role and Effects of the Present Invention Power Demand Prediction Prediction of power consumption by region and time of day Power consumption by past point in time, external temperature, humidity, day of the week information, etc. Power consumption for a specific future time period Efficient prediction possible by selecting only partial regions / indicators instead of the entire characteristic. Ensuring Robustness against Missing Sensors Traffic Flow Prediction Traffic volume prediction and congestion control for major roads Traffic volume, average speed, and weather information by road sensor Future traffic volume for each road Prediction by selecting only a subset out of hundreds of sensors. Reduction of computational load and improvement of inference accuracy Weather Change Prediction Prediction of environmental factors such as temperature, humidity, and atmospheric pressure Measurements by time of day for each weather station (temperature, wind speed, precipitation, etc.) Future temperature and humidity for a specific region, etc. Processing possible by selecting only a subset excluding missing regions. Adaptive Prediction Capability Patient Monitoring Prediction of patient vital signs and detection of anomalies Time flow of biological data such as heart rate, blood pressure, and oxygen saturation Prediction of heart rate and oxygen saturation for the next 1 to 2 hours Prediction possible even if some real-time biological signals are missing. Suitable for lightweight real-time inference Financial Asset Forecasting: Predicting specific asset prices or risks based on external indicators derived from historical prices, trading volume, interest rates, and news by asset, as well as future prices or volatility. High-precision forecasting is possible using meaningful subset combinations without handling all assets. Smart Energy Management System (SEMS) and Smart Appliance Interfacing Platform: Deriving optimal device operation schedules by predicting future power consumption and load based on composite sensor data from appliances and energy resources. Time-based appliance power consumption (refrigerators, washing machines, air conditioners, etc.), indoor and outdoor temperature and humidity, solar power generation, remaining battery capacity, and time factors (day of the week, time of day, etc.) Expected power usage, peak load times, and detailed load predictions for each device in 3 / 6 / 12-hour increments—efficiently performing forecasts by sampling only a portion of the overall characteristics.

[0130] In the fields of energy management and weather forecasting, an embodiment according to the present disclosure can be utilized for efficient forecasting of power demand and response to weather changes. When forecasting power demand, data such as past power usage, external temperature, and humidity are input to derive power usage for a certain period in the future; efficient forecasting can be achieved by selecting only a subset of regions or indicators instead of the entire set of indicators. In the case of forecasting weather changes, various environmental factors such as temperature and humidity are used as measurements, and since only a subset excluding missing regions can be selected and processed, adaptive weather forecasting becomes possible. This method ensures strong robustness, maintaining the model's forecasting performance even in real-world environments where some sensors fail or data is missing.

[0131] Furthermore, an embodiment according to the present disclosure can demonstrate excellent performance in complex systems such as traffic flow control and financial asset forecasting. In traffic flow forecasting, computational load can be reduced and inference accuracy improved by selecting only a portion of the data from hundreds of road sensors, and in the financial sector, risk can be predicted with high precision by utilizing asset-specific prices or news-based indicators. Unlike conventional methods that focus solely on a single optimal scheduling, this approach increases the probability of finding various optimal alternatives through multi-scenario planning. This approach can broaden the range of optimal choices available to users in response to ever-changing market or traffic conditions.

[0132] For example, the processor acquires multivariate financial time series data containing historical prices, trading volumes, interest rates, exchange rates, and news-based external economic indicators for each asset, with a total number of features D; generates multiple randomly extracted data by randomly extracting a subset of financial features S (1 < S < D) from the multivariate financial time series data; models the price fluctuation trend of individual assets by performing a time-axis attention operation on each of the multiple randomly extracted data within a shared transformer structure; extracts correlations between different assets within the same time interval or dependencies between specific economic indicators and asset prices by performing an attention operation between features based on the results of the time-axis attention operation; and calculates the average of individual prediction values ​​derived for each of the multiple randomly extracted data to determine a risk prediction result, including future price trends or volatility of a specific asset, as the final output value.

[0133] As another example, the processor acquires multivariate traffic time series data containing time-series traffic volume, average driving speed, and weather information collected from multiple road sensors, with a total number of features D, randomly extracts a subset of road sensors with a number of features S (1 < S < D) from the multivariate traffic time series data to generate multiple random extraction data, performs time-axis attention on each of the multiple random extraction data within a shared transformer structure to model the traffic flow pattern of individual roads, and performs attention operations between features on the multiple random extraction data to extract traffic volume interactions between adjacent roads or correlations between weather conditions and driving speed, and ensembles the predicted values ​​calculated for each of the multiple random extraction data to derive final predicted data including future traffic volume and congestion for each road.

[0134] In the fields of medical and smart home management, an embodiment according to the present disclosure can maximize the efficiency of real-time data processing. In patient monitoring systems, real-time prediction is possible through partial correlations of the remaining signals even if some vital signs, such as heart rate or blood pressure, are missing, which is particularly suitable for lightweight real-time inference devices. In smart energy management systems (SEMS), optimal device operation schedules are derived based on composite sensor data, such as the power consumption of home appliances and solar power generation; energy resources can be optimized by efficiently performing predictions by sampling only a portion of the total characteristics. Consequently, the present system can provide practical solutions for various real-world multivariate time series problems based on efficient feature attention costs and robustness in the event of missing features.

[0135] For example, the processor acquires multivariate energy time series data containing power consumption at specific times of multiple home appliances, indoor and outdoor temperature and humidity information, and external weather information, with a total number of features D; generates multiple randomly extracted data by randomly extracting a subset of energy features with a number of features S (1 < S < D) from the multivariate energy time series data; models the temporal dependency of individual energy features by performing a time-axis attention operation on each of the multiple randomly extracted data within a shared transformer structure; and extracts the correlation between the power consumption and temperature and humidity information of different home appliances within the same time interval by performing an attention operation between features based on the results of the time-axis attention operation, and calculates the average of the individual predicted values ​​calculated for each of the multiple randomly extracted data to determine a final output value including the expected power usage and optimal home appliance operation schedule for a certain time unit in the future.

[0136] FIG. 8 is a block diagram of a time series forecasting system according to one embodiment of the present disclosure.

[0137] Referring to FIG. 8, a time series prediction system (800) (the device may be referred to as a server or system) may include a transceiver (810), memory (820), a data storage unit (830), and a processor (850). However, not all components shown in FIG. 8 are essential components of the time series prediction system (800). The time series prediction system (800) may be implemented with more components than those shown in FIG. 8, or with fewer components than those shown in FIG. 8. Furthermore, the transceiver (810), memory (820), and processor (850) may be implemented in the form of a single chip.

[0138] In one embodiment, the transceiver (810) can communicate with a terminal or other electronic device connected to the time series prediction system (800) via wired or wireless connection.

[0139] Various types of data, such as programs and files, such as applications, can be installed and stored in the memory (820). The processor (850) may access and use the data stored in the memory (820) or store new data in the memory (820). Additionally, one or more instructions may be stored in the memory (820). The processor (850) may execute one or more instructions stored in the memory. The memory (820) may store information within the system. For example, the memory (820) may be a computer-readable medium, a volatile memory unit, or a non-volatile memory unit.

[0140] Functions related to artificial intelligence according to the present disclosure are operated through a processor and memory. The processor may be composed of one or more processors. In this case, the one or more processors may be general-purpose processors such as CPUs, APs, and DSPs (Digital Signal Processors), graphics-dedicated processors such as GPUs and VPUs (Vision Processing Units), or artificial intelligence-dedicated processors such as NPUs. The one or more processors control the processing of input data according to predefined operation rules or artificial intelligence models stored in memory. Alternatively, if the one or more processors are artificial intelligence-dedicated processors, the artificial intelligence-dedicated processors may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

[0141] In one embodiment, the data storage unit (830) may provide a large storage for the time series prediction system (800). For example, the data storage unit (830) may be a computer-readable medium. Alternatively, the data storage unit (830) may include a hard disk device, an optical disk device, a storage device shared over a network by a number of computing devices (e.g., a cloud storage device), or some other large-capacity storage device. The data storage unit (830) may include a trained neural network model (840).

[0142] The processor (850) controls the overall operation of the time series prediction system (800) and may include at least one processor, such as a CPU, GPU, etc. The processor (850) may control other components included in the time series prediction system (800) to perform operations for operating the time series prediction system (800). For example, the processor (850) may generate prediction data by processing an input sequence segmented along the time axis using a neural network by executing one or more instructions.

[0143] One embodiment of the present disclosure may also be implemented in the form of a recording medium comprising computer-executable instructions, such as program modules executed by a computer. A computer-readable medium may be any available medium accessible by a computer and includes both volatile and non-volatile media, and both removable and non-removable media. Additionally, a computer-readable medium may include both computer storage media and communication media. A computer storage medium includes both volatile and non-volatile, removable and non-removable media implemented by any method or technique for storing information, such as computer-readable instructions, data structures, program modules, or other data. A communication medium typically includes computer-readable instructions, data structures, or program modules and includes any information transmission medium.

[0144] The foregoing description of the present disclosure is for illustrative purposes only, and those skilled in the art will understand that modifications can be easily made to other specific forms without altering the technical spirit or essential features of the present invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form.

[0145] The scope of the present disclosure is defined by the claims set forth below rather than by the detailed description above, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts thereof should be interpreted as being included within the scope of the present disclosure.

Claims

1. In the system, One or more processors; and The system includes one or more memories that collectively store instructions that cause the system to perform operations when executed by the above-mentioned one or more processors, and the operations are: Operation of acquiring time series data; An operation to generate multiple randomly extracted data by performing multiple random extractions on the above time series data; An operation to perform an operation that models the temporal dependency of the plurality of randomly extracted data above; An operation to perform an operation to extract correlations between multiple features of the plurality of randomly extracted data; and A system comprising an operation to generate at least one prediction data based on at least one of the result of an operation to model the above-mentioned temporal dependency and the result of an operation to extract the correlation between the above-mentioned plurality of features.

2. In paragraph 1, the time series data includes D unit data, and The operation of generating the above plurality of randomly extracted data is, The operation of generating D unit data by dividing the above time series data based on the time axis; and The method includes an operation of generating multiple randomly extracted data by performing an operation of randomly selecting S unit data from the D unit data multiple times. A system in which S is a value greater than 1 and less than D.

3. In Paragraph 1, The operation of generating at least one prediction data above is, The method includes an operation to generate at least one prediction data by sequentially performing an operation to model the above temporal dependency and an operation to extract the correlation between the above plurality of features. The operation of performing an operation to extract correlations between the above plurality of features is, A system comprising an operation to extract correlations between a plurality of features with respect to the result of an operation modeling the above temporal dependency.

4. In claim 1, the operation of performing an operation to extract correlations between the plurality of features is, An operation to transform the array axis of the result of the operation modeling the above temporal dependency; and A system comprising an operation to extract correlations between a plurality of features based on a feature axis for data in which the array axis has been transformed.

5. In paragraph 1, the above operations are, The operation of generating a plurality of tokens by dividing the above plurality of randomly extracted data based on a time axis; and For each of the above plurality of tokens, the method further includes the operation of adding a location embedding representing temporal location information and characteristic location information, and A system in which the operation for modeling the above temporal dependency and the operation for extracting the correlation between the above plurality of features are performed based on a plurality of tokens to which the above position embeddings are added.

6. In paragraph 1, the operation of generating at least one prediction data is, The operation of calculating individual predicted values ​​for each of the above plurality of randomly extracted data; A system comprising the operation of generating at least one prediction data by calculating the average of individual prediction values ​​for each of the plurality of randomly extracted data.

7. In paragraph 1, the operation of generating at least one prediction data is, A system comprising the operation of generating at least one prediction data by inputting the result of an operation for modeling the temporal dependency and the result of an operation for extracting the correlation between the plurality of features into a Multi-Layer Perceptron (MLP) module.

8. A method for generating prediction data performed by at least one processor, Step of acquiring time series data; A step of generating multiple randomly extracted data by performing multiple random extractions on the above time series data; A step of performing an operation to model the temporal dependency of the plurality of randomly extracted data; A step of performing an operation to extract correlations between multiple features of the plurality of randomly extracted data; and A method comprising the step of generating at least one prediction data based on at least one of the result of an operation modeling the temporal dependency and the result of an operation extracting the correlation between the plurality of features.

9. In paragraph 8, the time series data includes D unit data, and The step of generating the above plurality of randomly extracted data is, A step of generating D unit data by dividing the above time series data based on the time axis; and The method includes the step of generating the plurality of randomly extracted data by performing the operation of randomly selecting S unit data from the D unit data multiple times. A method in which S is a value greater than 1 and less than D.

10. In Paragraph 8, The step of generating at least one prediction data above is, The method includes the step of generating at least one prediction data by sequentially performing the step of performing an operation to model the temporal dependency and the step of performing an operation to extract the correlation between the plurality of features. The step of performing an operation to extract correlations between the above plurality of features is: A method comprising the step of extracting correlations between the plurality of features with respect to the result of an operation modeling the above temporal dependency.

11. In claim 8, the step of performing an operation to extract correlations between the plurality of features is: A step of transforming the array axis of the result of the operation modeling the above temporal dependency; and A method comprising the step of performing an operation to extract correlations between a plurality of features based on a feature axis for data in which the array axis has been transformed.

12. In paragraph 8, the above method is, A step of generating a plurality of tokens by dividing the plurality of randomly extracted data based on a time axis; and For each of the plurality of tokens above, the method further includes the step of adding a location embedding representing temporal location information and characteristic location information, and A method in which the operation for modeling the above temporal dependency and the operation for extracting the correlation between the above plurality of features are performed based on a plurality of tokens to which the above position embeddings are added.

13. In paragraph 8, the step of generating at least one prediction data is, A step of calculating individual predicted values ​​for each of the above plurality of randomly extracted data; A method comprising the step of generating at least one prediction data by calculating the average of individual prediction values ​​for each of the plurality of randomly extracted data.

14. In claim 8, the step of generating at least one prediction data is, A method comprising the step of generating at least one prediction data by inputting the result of an operation modeling the temporal dependency and the result of an operation extracting the correlation between the plurality of features into a Multi-Layer Perceptron (MLP) module.

15. A computer-readable recording medium having a program recorded thereon for performing the method of any one of paragraphs 8 through 14 on a computer.