KR20210017342A - Time series prediction method and apparatus based on past prediction data - Google Patents

Abstract

Disclosed are a time-series prediction method and device based on past data. The time-series prediction method based on past data may comprise the steps of: obtaining real data measured in a prediction target environment; inputting the obtained real data into a previously learned time-series prediction model; obtaining primary prediction data predicted by the time-series prediction model; and obtaining the final prediction data by correcting the primary prediction data using past data extracted from a prediction data database (DB). Therefore, the time-series prediction method may obtain more accurate prediction data while using the previously learned time-series prediction model without re-learning.

Description

Time series prediction method and apparatus based on past data {TIME SERIES PREDICTION METHOD AND APPARATUS BASED ON PAST PREDICTION DATA}

The present invention relates to a time series prediction method and apparatus based on past data, and more particularly, by correcting a time series prediction result using only relevant past data without retraining a time series prediction model, a gradual or rapid data change If any, it relates to a technique for performing sophisticated time series prediction.

Time series refers to a sequence of data arranged at regular time intervals. At this time, time series prediction means creating a mathematical prediction model using given time series data, and predicting data that may occur in the future using the prediction model. In general, such time series prediction is widely used in engineering, scientific calculation, or stock price prediction in financial markets, but all other data measured dependent on time can be viewed as time series.

A prediction model for time series prediction may be generated by learning a correlation between input and output of known data. For example, by learning the correlation between the weather for the last three hours and the amount of electricity produced, a predictive model can be generated that predicts the amount of electricity produced according to a specific weather over the next hour.

The prediction model has high prediction accuracy when it is used in an environment similar to the environment in which the data used for training was collected. For example, a model for predicting power production generated using data collected in regions with distinct four seasons is not suitable for predicting the required power production in the equatorial region.

Meanwhile, the environment in which the predictive model is used may change gradually or rapidly. For example, the production capacity of a solar power plant to which a predictive model for predicting power production is applied gradually decreases over time due to aging, or the solar power plant's production capacity is caused by temporary snow or dust accumulation on the solar panel. This may decrease rapidly. Therefore, it is necessary to maintain the accuracy of the prediction model by appropriately correcting the real-time prediction result of the prediction model in consideration of changes in the environment in which the prediction model is used.

In order to solve this problem, conventionally, a method of fine-tuning a prediction model based on data generated in real time has been proposed, but many computing resources are required for further learning. In addition, in an environment where events occur intermittently for a short period of time, it is difficult to quickly and accurately additionally learn a prediction model, and a prediction model that is rapidly additionally learned using data generated by an intermittent event can generate inaccurate prediction data in a normal environment. It is likely to be derived.

An object of the present invention for solving the above problems is to provide a method for predicting a time series based on past data.

Another object of the present invention for solving the above problems is to provide an apparatus for predicting time series based on past data.

An aspect of the present invention for achieving the above object is to provide a time series prediction method based on past data.

The time series prediction method based on the past data includes: acquiring actual data observed in a prediction target environment, inputting the acquired actual data into a pre-trained time series prediction model, and a first-order prediction predicted by the time series prediction model. It may include obtaining data and correcting the first prediction data using past data extracted from a prediction data database (DB), thereby obtaining final prediction data.

The time series prediction method based on past data may further include storing the actual data in the prediction data DB.

The obtaining of the final prediction data may include extracting the past data according to a predetermined criterion from the prediction data DB.

The extracting of the past data may include extracting, from the predicted data DB, past predicted data or past actual data belonging to a range in which at least one of the first predicted data and numerical value, a current predicted time point and a source is adjacent. have.

In the extracting of the past data, when the prediction target environment is an environment in which there is a rapid environmental change, past data adjacent to the current prediction time point and belonging to a short time interval may be extracted.

The step of extracting the past data may further include pre-processing the data extracted from the prediction data DB according to a predetermined criterion.

The obtaining of the final predicted data may include correcting the past actual data to obtain corrected past actual data.

The obtaining of the final prediction data may include calculating a past prediction error using the past data, and obtaining error correction prediction data by correcting the primary prediction data using the calculated past prediction error. I can.

The obtaining of the final prediction data may include obtaining the final prediction data by additionally applying a Kalman filter technique to the corrected past actual data and the error correction prediction data.

The time series prediction model may be pre-trained in the same or similar environment as the prediction target environment.

Another aspect of the present invention for achieving the above object is to provide a time series prediction apparatus based on past data.

An apparatus for predicting a time series based on past data may include: at least one processor; And a memory for storing instructions instructing the at least one processor to perform at least one step.

The at least one step may include obtaining actual data observed in a prediction target environment; Inputting the acquired real data into a pre-trained time series prediction model; Obtaining primary prediction data predicted by the time series prediction model; And acquiring final prediction data by correcting the primary prediction data using past data extracted from a prediction data database (DB).

The at least one step may further include storing the actual data in the prediction data DB.

The obtaining of the final prediction data may include extracting the past data according to a predetermined criterion from the prediction data DB.

The extracting of the past data may include extracting, from the predicted data DB, past predicted data or past actual data belonging to a range in which at least one of the first predicted data and numerical value, a current predicted time point and a source is adjacent. have.

In the extracting of the past data, when the prediction target environment is an environment in which there is a rapid environmental change, past data adjacent to the current prediction time point and belonging to a short time interval may be extracted.

The step of extracting the past data may further include pre-processing the data extracted from the prediction data DB according to a predetermined criterion.

The obtaining of the final predicted data may include correcting the past actual data to obtain corrected past actual data.

The obtaining of the final prediction data may include calculating a past prediction error using the past data, and obtaining error correction prediction data by correcting the primary prediction data using the calculated past prediction error. I can.

The obtaining of the final prediction data may include obtaining the final prediction data by additionally applying a Kalman filter technique to the corrected past actual data and the error correction prediction data.

The time series prediction model may be pre-trained in the same or similar environment as the prediction target environment.

In the case of using the method and apparatus for predicting time series based on past data according to the present invention as described above, prediction data may be corrected in real time using only past prediction data even without additional learning of a prediction model.

In addition, there is an advantage in that prediction data can be quickly corrected and derived even in an environment in which unexpected intermittent events occur.

In addition, it is not dependent on the prediction model, and the prediction model generated by learning in a specific environment can be used directly in a similar environment.

1 is a conceptual diagram of a method and apparatus for predicting a time series based on past data according to an embodiment of the present invention.
2 is a conceptual diagram illustrating a method of extracting past data according to an embodiment of the present invention.
3 is a conceptual diagram of a first method for correcting primary prediction data using a prediction model based on past data according to an embodiment of the present invention.
4 is a conceptual diagram of a second method for correcting primary prediction data using a prediction model based on past data according to an embodiment of the present invention.
5 is a conceptual diagram of a third method of correcting primary prediction data using a prediction model based on past data according to an embodiment of the present invention.
6 is an exemplary diagram for explaining a process of generating prediction data for site 2 using a prediction data DB collected at site 1 and a time series prediction model learned at site 1;
7 is a graph showing the result of applying the time series prediction model learned in the first region to a second region belonging to a climate environment similar to that of the first region but having different power generation performance.
8 is a graph showing a result of applying a time series prediction model learned in a first region where power generation performance gradually decreases to a first region.
9 is a graph showing a result of applying a time series prediction model learned in a first region having a sudden environmental change to a first region.
10 is a representative flowchart of a method for predicting time series based on past data according to an embodiment of the present invention.
11 is a hardware configuration diagram of a time series prediction apparatus based on past data according to an embodiment of the present invention.

In the present invention, various modifications may be made and various embodiments may be provided, and specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to a specific embodiment, it is to be understood to include all changes, equivalents, and substitutes included in the spirit and scope of the present invention. In describing each drawing, similar reference numerals have been used for similar elements.

Terms such as first, second, A, and B may be used to describe various elements, but the elements should not be limited by the terms. These terms are used only for the purpose of distinguishing one component from another component. For example, without departing from the scope of the present invention, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element. The term and/or includes a combination of a plurality of related listed items or any of a plurality of related listed items.

When a component is referred to as being "connected" or "connected" to another component, it is understood that it may be directly connected or connected to the other component, but other components may exist in the middle. Should be. On the other hand, when a component is referred to as being "directly connected" or "directly connected" to another component, it should be understood that there is no other component in the middle.

The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present application, terms such as "comprise" or "have" are intended to designate the presence of features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, but one or more other features. It is to be understood that the presence or addition of elements or numbers, steps, actions, components, parts, or combinations thereof, does not preclude in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms as defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in this application. Does not.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a conceptual diagram of a method and apparatus for predicting time series based on past prediction data according to an embodiment of the present invention.

Referring to FIG. 1, a time series prediction method and apparatus based on past prediction data according to an embodiment of the present invention includes a time series prediction model 10 that is pre-trained and generated, and prediction data predicted by a time series prediction model 10. The final predicted data is output by correcting the primary predicted data output by the time series predictive model 10 based on the past data obtained from the predicted data DB 20 and/or the predicted data DB 20 storing actual data and It may be configured with a predictive data adjuster 30.

The time series prediction model 10 is a mathematical model for predicting result data that may occur in the future by learning time series data, and may include an autoregressive (AR) model, an integrated (I) model, and a moving average (MA) model. This time series prediction model (10) is based on machine learning algorithms such as long short-term memory (LSTM) using various learning methods including machine learning and long short-term memory (LSTM). Thus, it can be generated by learning various actual data (or observed or measured data). In this case, the actual data may be data observed or measured in the environment to be predicted (05). For example, when the environment to be predicted 05 is a power plant, the actual data may be an amount of power produced by the power plant, weather, and the like.

Accordingly, the time series prediction model 10 may receive such various types of real data as input data and output primary prediction data that predicts a specific time or time point in the future from the present time point. In this case, the input data may have various data types. For example, the input data for the time series prediction model 10 may be any n-dimensional time series data or time series data measured or collected during the last K time units (hours, minutes, days, etc.). Further, the primary prediction data corresponding to the output data of the time series prediction model 10 may be data that predicts a result for a specific time unit in the future from the current point in time.

Specifically, for example, the time series prediction model 10 may predict the amount of generation expected to be produced for an hour from the current point in time by using the amount of generation according to the weather for the last 3 hours as input data.

In an embodiment of the present invention, the time series prediction model 10 is not limited to a specific form or method and is not interpreted, and any time series prediction model implemented in various formats and algorithms may be included.

In addition, the first prediction data predicted for the prediction target environment 05 by the time series prediction model 10 and the actual data obtained by measuring or observing the prediction target environment 05 are separately constructed prediction data DB 20 ) Can be stored and managed. Accordingly, the prediction data DB 20 accumulates and stores the results primarily predicted for a specific time or period in the future by the time series prediction model 10 and actual data at the predicted time or period. The past data required by (30) can be provided.

For example, data stored and managed in the prediction data DB 20 may be as shown in Table 1 below.

Referring to Table 1, in the prediction data DB 20, a source of stored data, a time period in which observation or prediction is performed (Time index), and prediction data according to a time period (Forecast value, primary prediction data) May be, or may be referred to as past prediction data), actual data according to time intervals (True value, data actually collected in the prediction target environment or may be referred to as past actual data), various additional data on time intervals (Additional data1, Additional data2), etc. may be stored as fields. The additional data is data other than the predicted target indicator, and when predicting the amount of power generation, the additional data may be weather, age, etc. other than the amount of power generation.

In addition, the source may be an identification symbol indicating the prediction target environment from which the prediction data or actual data is obtained. For example, data having a prediction target environment to which the time series prediction model is to be applied or an environment similar to the prediction target environment as a source may be stored in the prediction data DB 20.

The prediction data adjuster 30 may output final prediction data by correcting the primary prediction data output by the time series prediction model based on past data provided from the prediction data DB. Here, the past data may be past prediction data that the time series prediction model 10 predicts for a future point or time from a specific point in the past, or past actual data measured, collected, or observed at a specific point in the past. In addition, the past data may be data having high relevance to primary prediction data predicted by a time series prediction model among past prediction data or past actual data.

Meanwhile, the time series prediction model 10 and the prediction data adjuster 30 may be implemented as software modules or instructions that perform the above-described functions or operations. In addition, the prediction data DB 20 may be implemented as a hardware storage, a data server, a cloud server, or the like that stores and manages the aforementioned data.

2 is a conceptual diagram illustrating a method of extracting past prediction data according to an embodiment of the present invention.

As described in FIG. 1, the prediction data adjuster 30 according to an embodiment of the present invention extracts past data stored in the prediction data DB 20, and uses the extracted past data to make a first-order prediction of a time series prediction model. Data can be corrected.

Here, the past data extracted from the prediction data DB 20 may be data similar to the first prediction data according to the time series prediction model. For example, the past data may be data that falls within a predetermined reference range value with the primary prediction data or within a predetermined ratio range with respect to the primary prediction data.

In addition, the past data may be data belonging to a time interval determined based on a current time point from which primary prediction data according to a time series prediction model is derived. For example, the past data may be a time section in which the current time point from which the primary prediction data is derived belongs or data located within a certain time range from the current time point. More specifically, for example, in the case of predicting the amount of electricity produced at 12 pm, past actual data collected at 12 pm or past prediction data predicted at 12 pm may be extracted as past data.

Also, the past data may be data determined based on a prediction target environment to which the time series prediction model is currently applied. For example, the past data may be data that has been collected or predicted in an environment similar to the environment to be predicted.

The past data may include all fields (refer to fields according to Table 1) or some fields included in the prediction data DB 20. For example, past data may include time intervals, past prediction data, and past actual data.

Also, the past data may be randomly selected data from past data included in the prediction data DB 20.

The number of past data extracted from the prediction data DB 20 may be dynamically determined. For example, it is possible to dynamically determine the number of past data in consideration of a recent prediction result. Specifically, when it is determined that a specific event has occurred, the number of past data may be increased or decreased. Also, the number of past data may be determined based on a prediction target environment to which the time series prediction model is applied. For example, when the environment to be predicted corresponds to an environment in which there are few sudden events, a large number of past data may be input to the preprocessor 40. In addition, when the prediction target environment corresponds to an environment in which sudden events are frequent, a small number of past data (or data belonging to a very close time interval based on the current time point) may be input to the preprocessor 40. As described above, when prediction is made in an environment where sudden events are frequent, rapid changes due to sudden events can be quickly reflected in prediction by using past data for a very short time.

Meanwhile, the past data determined as described above may be used to correct the primary prediction data according to FIG. 1 as it is in the prediction data adjuster 30, or preprocessed through the preprocessor 40 as in FIG. Can be used in (30). For example, the preprocessor 40 may sort past data according to a certain criterion, select a part of the sorted data, and provide it to the predictive data adjuster 30. In addition, the preprocessor 40 may simultaneously extract past actual data belonging to different numbers of time intervals from the predicted data DB 20, and preprocess the extracted past actual data to derive one past actual data. . For example, it is possible to extract past actual data belonging to the past 1000 time intervals, extract past actual data belonging to the past 10 time intervals at the same time, and preprocess the extracted data to derive one past actual data. have. One past actual data derived here may be transmitted to the predicted data adjuster 30.

Past data preprocessed or determined through the above-described process may be used to correct primary prediction data according to a time series prediction model. Hereinafter, a method of correcting primary prediction data using past data will be described in detail.

3 is a conceptual diagram of a first method for correcting primary prediction data using a prediction model based on past data according to an embodiment of the present invention. 4 is a conceptual diagram of a second method for correcting primary prediction data using a prediction model based on past data according to an embodiment of the present invention. 5 is a conceptual diagram of a third method of correcting primary prediction data using a prediction model based on past data according to an embodiment of the present invention.

Referring again to FIG. 1, final prediction data according to an embodiment of the present invention may be derived by correcting primary prediction data according to a time series prediction model using past data obtained from a prediction data DB. In this case, as a method for the prediction data adjuster 30 to derive the final prediction data, a first method according to FIG. 3, a second method according to FIG. 4, and a third method according to FIG. 5 may be considered.

First, referring to FIG. 3, corrected past actual data may be derived by correcting the past actual data extracted from the prediction data DB, and final predicted data may be derived from the corrected past actual data.

For example, the corrected past actual data may be determined as in Equation 1 below.

Referring to Equation 1, the corrected past actual data can be derived by adding a correction value (the parenthesis portion of Equation 1) to the past actual data. Here, the correction value may be determined by an operation using primary prediction data and past prediction data. In addition, the coefficients α, β, and γ for each data in Equation 1 are weights and may be changed according to a user's setting.

Accordingly, when the corrected past actual data is derived according to Equation 1, the final predicted data can be derived using the past actual data. For example, an average value or a weighted average value of corrected past actual data may be calculated and used as final prediction data.

Referring to FIG. 4, final prediction data may be derived by calculating a past prediction error from past data extracted from a prediction data DB, and correcting primary prediction data using the calculated past prediction error.

First, the past prediction error may be determined as in Equation 2 below.

Referring to Equation 2, the past prediction error may be determined by a difference value between past actual data and past prediction data. In this case, weights α and β may be assigned to the past actual data and past predicted data, respectively. According to Equation 2, past prediction errors may be calculated in a time interval that is the same as or similar to the current time point, and an average or weighted average of the calculated past prediction errors may be calculated and finally used as the past prediction errors.

Next, the first prediction data may be corrected by using the past prediction error as shown in Equation 3 below.

Referring to Equation 3, error correction prediction data may be derived by adding past prediction errors to primary prediction data, and such error correction prediction data may be used as final prediction data.

Referring to FIG. 5, final predicted data may be determined by additionally applying a Kalman filter technique to corrected past actual data according to Equation 1 and error correction predicted data according to Equation 3.

In this case, since the Kalman filter technique multiplies two Gaussian functions to derive one Gaussian function, it is necessary to calculate a standard deviation for each of the corrected past actual data and the error-corrected predicted data according to Equation 3.

Here, the standard deviation of the corrected past actual data according to Equation 1 may be provided from the time series prediction model, or may be obtained by transforming the prediction interval provided by the time series prediction model into a standard deviation, and the distribution of the past prediction data It can also be calculated directly based on.

In addition, the standard deviation of the error correction predicted data according to Equation 3 may be a value corresponding to a specific percentile with respect to the difference between the past actual data and the past predicted data, or the past predicted data and the actual predicted data. It may be the number of maximum difference values, or may be directly calculated based on the distribution of actual data in the past.

When the standard deviation is calculated for each of the corrected past actual data according to Equation 1 and the error correction predicted data according to Equation 3, the calculated standard deviations and corrected past actual data according to Equation 1 and Equation 3 are One average and standard deviation can be derived by using the error-corrected prediction data. At this time, one average value finally derived may be used as final prediction data according to the Kalman filter technique.

Referring to FIG. 5, the predicted data adjuster 30 according to FIG. 1 calculates corrected past actual data and error correction predicted data using past actual data and past predicted data, and standard deviation for each of the calculated data. After calculating to, an average value according to the Kalman filter 31 is derived and used as final prediction data can be confirmed.

Although not shown in FIGS. 3 to 5, the criteria and number of extracting past prediction data or past actual data extracted from the past prediction DB 20 should be interpreted as being applicable to the methods according to FIG. 2.

6 is an exemplary diagram for explaining a process of generating prediction data for site 2 using a prediction data DB collected at site 1 and a time series prediction model learned at site 1;

In FIG. 6, the prediction data DB 20 may be composed of data collected and predicted in the first area (site 1), and the time series prediction model 10 may also be a model learned in the first area (site 1). have. On this premise, a case of performing time series prediction on the second site may be considered.

For example, the first area is a solar power plant with a capacity of 99 KWp installed in the Yeongam area, and the second area is installed in Jangheung, which has a close distance from the first area and has a similar climate, and has the same capacity as the first area (99 KWp) or It may be a solar power plant with a capacity (200 KWp) different from that of 1 area.

In this way, if you want to generate prediction data for the second area (site 2) by using the database and model collected and predicted in the first area, the predictive data adjuster will initially use only the data collected in the first area. Since the predicted data is adjusted, characteristics according to the second region may not be accurately reflected.

However, as the prediction continues, the prediction data DB may include past data collected in the second area. That is, the prediction data DB may include both past data collected in the first area and past data collected in the second area. Accordingly, as time passes, the number of historical data collected in the second area increases, and thus the prediction accuracy for the second area may be further increased.

In addition, when sufficient past data is collected in the second area, the first prediction data may be corrected using only the past data collected in the second area. That is, by using only past data collected in the second area, accuracy of final prediction data may be improved through correction of the primary prediction data.

Unlike the above process, the prediction data DB collected in the first area and the time series prediction model learned in the first area are not used, but the prediction data DB collected and learned in the second area and the time series prediction model are also used. It is possible.

7 is a graph showing the result of applying the time series prediction model learned in the first region to a second region belonging to a climate environment similar to that of the first region but having different power generation performance.

Referring to FIG. 7, a time series prediction model learned based on a first region (or a first power plant) is a second region (or a second power plant) that belongs to a climate environment similar to that of the first region, but has a low production performance of about 30%. ), you can check the results.

In FIG. 7,'TRUE' is the actual power production per unit time of,'Not adjusted' is the first prediction data of the time series prediction model learned in the first region as it is, and'Adjusted' is the first prediction data of the present invention. This shows the final predicted data derived by correcting using past data according to an embodiment of

When the time-series prediction model learned in the first region is applied to the second region as it is, it is observed that a large error from the actual power generation (TURE) occurs due to the difference in production performance, even if a similar climate environment is shown as the result of not adjusted can see.

On the other hand, in the case of correcting the primary prediction data of the time series prediction model based on the past data of the second area according to an embodiment of the present invention, there is an error to some extent when the data is insufficient (up to a time index value of about 5). However, it can be seen that as the data of the second area is gradually secured, the final predicted data (Adjusted) does not show a large error with the actual power generation amount (TURE).

8 is a graph showing a result of applying a time series prediction model learned in a first region where power generation performance gradually decreases to a first region.

Referring to FIG. 8, a prediction result can be confirmed for a case in which the region from which the time series prediction model is trained and the region to which the time series prediction model is trained and applied is the same, but the environment of the region is gradually changed (or the production performance is reduced by up to 30% due to aging).

In FIG. 8,'TRUE' is the actual power production per unit time of,'Not adjusted' is the first prediction data of the time series prediction model learned in the first region as it is, and'Adjusted' is the first prediction data of the present invention. This shows the final predicted data derived by correcting using past data according to an embodiment of

First, it can be seen that in the case of uncorrected primary prediction data, an error with the actual data TRUE increases over time.

On the other hand, the final prediction data derived by performing correction of the primary prediction data according to an embodiment of the present invention can reflect even deterioration over time, and thus it can be seen that a prediction error hardly occurs.

9 is a graph showing a result of applying a time series prediction model learned in a first region having a sudden environmental change to a first region.

In FIG. 9,'TRUE' is the actual power production per unit time of,'Not adjusted' is the first prediction data of the time series prediction model learned in the first region as it is, and'Adjusted' is the first prediction data of the present invention. This shows the final predicted data derived by correcting using past data according to an embodiment of

For example, in FIG. 9, prediction performance for a case in which power generation performance is impaired by sudden occurrence such as snow or shadow in a solar power plant according to the first region can be confirmed.

Referring to FIG. 9, it can be seen that the actual power generation (TURE) decreases due to a sudden external factor at the time points corresponding to 11 to 40 of the time index, and the power generation performance decreases by about 30%.

Looking at the case that the time series prediction result is not corrected (Not Adjusted) when the performance deteriorates due to an abrupt factor, it can be seen that the prediction error with the actual power generation is maintained large and the prediction error decreases again only after the power generation performance is restored .

However, it can be seen that the final data corrected by using the past data according to an embodiment of the present invention has a large error similar to the case not corrected in the early stage of sudden performance degradation, but gradually decreases the error. . In addition, it can be seen that even after the power generation performance is restored, prediction results are derived without significant errors. As described above, when a sudden change occurs in the environment to be predicted, the past data used for correction may be used as data for the recent short term or in a small number so that such change may be quickly and accurately reflected as a prediction result.

10 is a representative flowchart of a method for predicting time series based on past data according to an embodiment of the present invention.

Referring to FIG. 10, the method of predicting a time series based on past data includes: obtaining actual data observed in a prediction target environment (S100), inputting the obtained actual data into a pre-trained time series prediction model (S110), Acquiring the first prediction data predicted by the time series prediction model (S120) and correcting the first prediction data using past data extracted from a prediction data DB (database), thereby obtaining final prediction data ( S130) may be included.

The time series prediction method based on past data may further include storing the actual data in the prediction data DB.

Acquiring the final prediction data (S130) may include extracting the past data according to a predetermined criterion from the prediction data DB.

The extracting of the past data may include extracting, from the predicted data DB, past predicted data or past actual data belonging to a range in which at least one of the first predicted data and numerical value, a current predicted time point and a source is adjacent. have.

In the extracting of the past data, when the prediction target environment is an environment in which there is a rapid environmental change, past data that is adjacent to the current prediction time and belongs to a short time interval may be extracted.

The step of extracting the past data may further include pre-processing the data extracted from the prediction data DB according to a predetermined criterion.

Acquiring the final predicted data (S130) may include correcting the past actual data to obtain corrected past actual data.

In the step of obtaining the final prediction data (S130), obtaining error correction prediction data by calculating a past prediction error using the past data and correcting the primary prediction data using the calculated past prediction error It may include.

The obtaining of the final prediction data (S130) may include obtaining the final prediction data by additionally applying a Kalman filter technique to the corrected past actual data and the error correction prediction data. have.

The time series prediction model may be pre-trained in the same or similar environment as the prediction target environment.

11 is a hardware configuration diagram of a time series prediction apparatus based on past data according to an embodiment of the present invention.

Referring to FIG. 11, the apparatus 100 for predicting a time series based on past data includes at least one processor 110; And a memory 120 for storing instructions instructing the at least one processor to perform at least one step.

Here, the at least one processor 110 means a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated processor in which the methods according to the embodiments of the present invention are performed. I can. Each of the memory 120 and the storage device 160 may be configured with at least one of a volatile storage medium and a nonvolatile storage medium. For example, the memory 120 may be composed of at least one of a read only memory (ROM) and a random access memory (RAM).

In addition, the time series prediction apparatus 100 based on past data may include a transceiver 130 that performs communication through a wireless network. In addition, the time series prediction apparatus 100 based on past data may further include an input interface device 140, an output interface device 150, and a storage device 160. Each component included in the time series prediction apparatus 100 based on past data may be connected by a bus 170 to communicate with each other.

The at least one step may include obtaining actual data observed in a prediction target environment; Inputting the acquired real data into a pre-trained time series prediction model; Obtaining primary prediction data predicted by the time series prediction model; And obtaining final prediction data by correcting the primary prediction data using past data extracted from a prediction data database (DB).

The at least one step may further include storing the actual data in the prediction data DB.

The obtaining of the final prediction data may include extracting the past data according to a predetermined criterion from the prediction data DB.

The extracting of the past data may include extracting, from the predicted data DB, past predicted data or past actual data belonging to a range in which at least one of the first predicted data and numerical value, a current predicted time point and a source is adjacent. have.

In the extracting of the past data, when the prediction target environment is an environment in which there is a rapid environmental change, past data that is adjacent to the current prediction time and belongs to a short time interval may be extracted.

The step of extracting the past data may further include pre-processing the data extracted from the prediction data DB according to a predetermined criterion.

The obtaining of the final predicted data may include correcting the past actual data to obtain corrected past actual data.

The obtaining of the final prediction data may include calculating a past prediction error using the past data, and obtaining error correction prediction data by correcting the primary prediction data using the calculated past prediction error. I can.

The obtaining of the final prediction data may include obtaining the final prediction data by additionally applying a Kalman filter technique to the corrected past actual data and the error correction prediction data.

The time series prediction model may be pre-trained in the same or similar environment as the prediction target environment.

Examples of the time series prediction apparatus 100 based on past data include, for example, a communicable desktop computer, a laptop computer, a notebook, a smart phone, a tablet PC, and Mobile phone, smart watch, smart glass, e-book reader, portable multimedia player (PMP), portable game console, navigation device, digital camera, DMB (digital multimedia broadcasting) player, digital audio recorder, digital audio player, digital video recorder, digital video player, PDA (Personal Digital Assistant), etc. I can.

The methods according to the present invention may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like alone or in combination. The program instructions recorded on the computer-readable medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in computer software.

Examples of computer-readable media may include hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions may include high-level language codes that can be executed by a computer using an interpreter as well as machine language codes such as those produced by a compiler. The above-described hardware device may be configured to operate as at least one software module to perform the operation of the present invention, and vice versa.

In addition, the above-described method or apparatus may be implemented by combining all or part of its configuration or function, or may be implemented separately.

Although the above has been described with reference to preferred embodiments of the present invention, those skilled in the art will variously modify and change the present invention within the scope not departing from the spirit and scope of the present invention described in the following claims. You will understand that you can do it.

Claims (20)

A time series prediction method based on historical data,
Acquiring actual data observed in the environment to be predicted;
Inputting the acquired real data into a pre-trained time series prediction model;
Obtaining primary prediction data predicted by the time series prediction model; And
Comprising the step of obtaining final predicted data by correcting the primary predicted data using past data extracted from a predicted data database (DB). In claim 1,
The method of predicting a time series based on past data, further comprising storing the actual data in the prediction data DB. In claim 1,
Obtaining the final prediction data,
And extracting the past data according to a predetermined criterion from the prediction data DB. In claim 3,
Extracting the past data,
And extracting, from the prediction data DB, past prediction data or past actual data belonging to a range in which at least one of the first prediction data and numerical values, a current prediction time point, and a source are adjacent to each other. In claim 3,
Extracting the past data,
When the prediction target environment is an environment in which there is a sudden change in environment, past data adjacent to the current prediction time and belonging to a short time interval are extracted. In claim 3,
Extracting the past data,
Pre-processing the data extracted from the prediction data DB according to a predetermined criterion, a time series prediction method based on past data. In claim 4,
Obtaining the final prediction data,
Comprising the step of obtaining corrected past actual data by correcting the past actual data, a time series prediction method based on past data. In claim 7,
Obtaining the final prediction data,
And obtaining error correction prediction data by calculating a past prediction error using the past data and correcting the primary prediction data using the calculated past prediction error. In claim 8,
Obtaining the final prediction data,
And obtaining the final prediction data by additionally applying a Kalman filter technique to the corrected past actual data and the error correction prediction data. In claim 1,
The time series prediction model,
Time series prediction method based on past data, which is learned in advance in the same or similar environment as the prediction target environment. As a time series prediction device based on past data,
At least one processor; And
Including a memory (memory) for storing instructions (instructions) instructing the at least one processor to perform at least one step,
The at least one step,
Acquiring actual data observed in the environment to be predicted;
Inputting the acquired real data into a pre-trained time series prediction model;
Obtaining primary prediction data predicted by the time series prediction model; And
Comprising the step of obtaining final prediction data by correcting the primary prediction data using past data extracted from a prediction data DB (database). In claim 11,
The at least one step,
The time series prediction apparatus based on past data, further comprising storing the actual data in the prediction data DB. In claim 11,
Obtaining the final prediction data,
And extracting the past data according to a predetermined criterion from the prediction data DB. In claim 13,
Extracting the past data,
And extracting, from the prediction data DB, past prediction data or past actual data belonging to a range in which at least one of the first prediction data and numerical values, a current prediction time point, and a source are adjacent to each other. In claim 13,
Extracting the past data,
When the prediction target environment is an environment in which there is a rapid environmental change, the time series prediction apparatus based on past data extracts past data that is adjacent to the current prediction time and belongs to a short time interval. In claim 13,
Extracting the past data,
Pre-processing the data extracted from the prediction data DB according to a predetermined criterion, a time series prediction apparatus based on past data. In claim 14,
Obtaining the final prediction data,
Comprising the step of obtaining corrected past actual data by correcting the past actual data, a time series prediction apparatus based on past data. In claim 17,
Obtaining the final prediction data,
And obtaining error correction prediction data by calculating a past prediction error using the past data and correcting the primary prediction data using the calculated past prediction error. In claim 18,
Obtaining the final prediction data,
And obtaining the final prediction data by additionally applying a Kalman filter technique to the corrected past actual data and the error correction prediction data. In claim 11,
The time series prediction model,
A time series prediction apparatus based on past data that is learned in advance in an environment that is the same as or similar to the prediction target environment.

Images