Spike prediction device and spike prediction method

Abstract

A spike prediction device (100) comprises: a spike extraction unit (111) that extracts a time at which a spike has occurred, said spike being included in time-series data spanning one or more periods having a prescribed length; and a prediction unit (see the spike occurrence prediction unit (117)) that predicts, on the basis of a time at which a spike has occurred and time-series data of a period preceding a prediction target period which is a spike occurrence prediction target, a time in said prediction target period at which a spike will occur.

Description

Spike prediction device and spike prediction method

[0001] The present invention relates to a spike prediction device and a spike prediction method for predicting spikes contained in time-series data.

[0002] In wholesale electricity trading, power generation companies and retailers trade electricity for the next day or the current day, and electricity prices change constantly. For power generation companies and retailers, it is important to predict electricity prices in order to determine the operation of power plants, respond to the risk of price hikes, and calculate income and expenditures. The electricity demand forecasting system described in Patent Literature 1 divides an electricity supply area into multiple subareas based on human movement attributes that are highly correlated with fluctuations in electricity demand, and predicts future electricity demand for each subarea based on historical electricity demand data. Demand forecasts by such an electricity demand forecasting system can be used to predict trading prices.

[0003] Japanese Patent Application Laid-Open No. 2022-018848

[0004] Electricity prices, which are time-series data, may contain spike-like fluctuations (also simply referred to as spikes). Spikes are thought to occur due to factors other than periodic fluctuations such as days, weeks, or seasons, and price predictions that include spikes will result in reduced accuracy. For this reason, it is desirable to predict the occurrence and price of spikes separately from periodic fluctuations. This applies not only to electricity prices, but also to other predictions (regressions) that include spikes. The present invention has been made in light of this background, and aims to provide a spike prediction device and a spike prediction method that enable the prediction of spikes contained in time-series data.

[0005] In order to solve the above-mentioned problems, the spike prediction device of the present invention comprises a spike extraction unit that extracts the time at which a spike has occurred included in time-series data spanning one or more periods of a predetermined length, and a prediction unit that predicts the time at which the spike will occur in a prediction target period, which is the period for which spike occurrence is to be predicted, based on the time-series data and the time at which the spike occurred in the period prior to the prediction target period.

[0006] According to the present invention, it is possible to provide a spike prediction device and a spike prediction method that enable prediction of spikes contained in time-series data. Problems, configurations, and effects other than those described above will become apparent from the description of the following embodiments.

[0007] FIG. 1 is a functional block diagram of a spike prediction device according to the present embodiment. FIG. 2 is a graph showing power price data according to the present embodiment. FIG. 3 is a graph of a smoothing result obtained by smoothing power price data according to the present embodiment. FIG. 4 is a graph of a smoothed difference obtained by subtracting the smoothed result from power price data according to the present embodiment. FIG. 5 is power price data in which spikes have been removed from the smoothed difference of power price data according to the present embodiment. FIG. 6 is a flowchart of a spike extraction process according to the present embodiment. FIG. 7 is a flowchart of a spike occurrence prediction model generation process according to the present embodiment. FIG. 8 is a flowchart of a spike occurrence prediction process according to the present embodiment. FIG. 9 is a block diagram of a computer according to the present embodiment.

[0008] A spike prediction device according to an embodiment of the present invention will be described below. The spike prediction device decomposes time series data after spike removal into a trend component, a periodic component, and a residual component. The spike prediction device adjusts the decomposition parameters so as to optimize the component decomposition index. A spike in time series data is data with an instantaneous large or small value (at a certain time), which exhibits abrupt and large fluctuations in value and is considered abnormal.

[0009] Next, the spike prediction device extracts spikes contained in the trend difference, which is obtained by subtracting the trend component from the time-series data, and obtains the timing (time) and value (price) of the spike.The spike prediction device then predicts the timing and value of the spike using machine learning technology based on the obtained timing and value.

[0010] In this way, the spike prediction device can predict spikes with high accuracy. In the following explanation, the electricity price will be used as an example. The timing refers to the time of day. The spike prediction device predicts the time of occurrence and price of a spike on the prediction target day (the day on which a spike is desired to be predicted) based on price data (time series data) up to the day before the prediction target day.

[0011] <Configuration of the Spike Prediction Device> Fig. 1 is a functional block diagram of a spike prediction device 100 according to this embodiment. The spike prediction device 100 is a computer 900 (see Fig. 9 described below) and includes a control unit 110, a storage unit 130, and an input / output unit 180. User interface devices such as a display, keyboard, and mouse are connected to the input / output unit 180. The input / output unit 180 may include a communication device, enabling data transmission and reception with a device such as a server that provides electricity market prices. A media drive may also be connected to the input / output unit 180, enabling data exchange using a recording medium.

[0012] <Spike Prediction Device: Storage Unit> The storage unit 130 is configured to include storage devices such as a ROM (Read Only Memory), a RAM (Random Access Memory), and an SSD (Solid State Drive). The storage unit 130 stores a price database 140, a spike database 150, a component database 160, search parameters 131, a prediction model 132, and a program 138. The program 138 includes descriptions of a spike extraction process (see FIG. 6 described below), a spike occurrence prediction model generation process (see FIG. 7 described below), a spike occurrence prediction process (see FIG. 8 described below), and the like.

[0013] The price database 140 stores electricity prices (also referred to as electricity price data), which are time-series data. The spike database 150 stores the occurrence dates and times of spikes 331, 332 (see FIG. 4 , described later), which are sudden (instantaneous) price fluctuations included in the electricity price data, and the prices (electricity prices) at those times. The component database 160 stores trend components, periodic components, and residual components, which are the results of component decomposition of smoothed differences after spike removal, which will be described later. The component database 160 also stores component strength indices (also simply referred to as indices) of the trend components and periodic components.

[0014] The search parameters 131 store candidate parameters to be referenced when component-decomposing the electricity price data. The parameters include parameters used for smoothing by the smoothing unit 112 (described later) and parameters used for component-decomposition by the component decomposition unit 113. The prediction model 132 is a model that predicts the time (timing) of spike occurrence and price. The prediction model 132 includes a spike occurrence prediction model that predicts the timing of spike occurrence and a spike price prediction model that predicts the price.

[0015] <Spike Prediction Device: Control Unit> The control unit 110 is configured to include a CPU (Central Processing Unit) and is equipped with a spike extraction unit 111, a smoothing unit 112, a component decomposition unit 113, a component intensity evaluation unit 114, a trend difference calculation unit 115, a prediction model generation unit 116, a spike occurrence prediction unit 117, and a spike price prediction unit 118. The control unit 110 may be configured to include a GPU (Graphics Processing Unit), an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), etc.

[0016] The spike extraction unit 111 extracts and removes spikes included in the electricity price data. More specifically, the spike extraction unit 111 extracts and removes spikes included in the smoothed difference (see FIG. 4 ) calculated by the smoothing unit 112 described below. The spike extraction unit 111 extracts outliers in the interquartile range method, for example, as spikes. The spike extraction unit 111 replaces the price at the time of the extracted spike with a price linearly interpolated using the prices before and after it, thereby generating spike-removed time series data.

[0017] The spike extraction unit 111 also extracts spikes included in the trend difference calculated by the trend difference calculation unit 115 (described later), and stores the occurrence time and price (trend difference value) in the spike database 150. The extraction method is the same as the method used to extract outliers included in the smoothed difference.

[0018] The smoothing unit 112 smooths the electricity price data and calculates a smoothed difference by subtracting the smoothed result from the electricity price data. Known smoothing methods include moving average, Savitzky-Golay method, and Gaussian filter. The parameters used for smoothing are included in the search parameters 131. The processing of the spike extraction unit 111 and the smoothing unit 112 will be described below with reference to FIGS. 2 to 5.

[0019] FIG. 2 is a graph showing power price data according to this embodiment. The horizontal axis of the graph represents time, and the vertical axis represents price (power price). FIG. 3 is a graph of the smoothing result obtained by smoothing the power price data according to this embodiment. FIG. 4 is a graph of the smoothed difference obtained by subtracting the smoothed result (see FIG. 3) from the power price data according to this embodiment (see FIG. 2). The spike extraction unit 111 extracts and removes spikes 331 and 332 included in this smoothed difference. FIG. 5 is power price data in which the spikes 331 and 332 have been removed from the smoothed difference of the power price data according to this embodiment (see FIG. 4). The power price data obtained by adding the smoothing result (see FIG. 3) to this power price data has the same value (power price) except for the times of the spikes 331 and 332.

[0020] As described above, the spike prediction device 100 includes a spike extraction unit 111 that extracts the time points at which spikes occur in time-series data (electricity price data) spanning one or more periods of a predetermined length (one day). The spike extraction unit 111 calculates smoothed difference time-series data after spike removal (smoothed difference after spike removal, see FIG. 5 ), which is data obtained by removing spikes from smoothed difference data (smoothed difference), which is data obtained by subtracting smoothed time-series data from the time-series data. The spike extraction unit 111 also extracts the time points in the periods at which spikes occur in trend difference data (described below).

[0021] Returning to FIG. 1 , we will continue to explain the control unit 110. The component decomposition unit 113 performs component decomposition, which breaks down the spike-removed smoothed difference (see FIG. 5 ) into a trend component, a periodic component, and a residual component. The trend component represents the average trend of the time-series data (in this embodiment, the spike-removed smoothed difference). The periodic component represents the trend of the value (price) over a specified period. The residual component represents the trend of the value obtained by subtracting the trend component and periodic component from the time-series data. The component decomposition unit 113 performs component decomposition of the spike-removed smoothed difference using, for example, a technique called seasonal classical decomposition. The period is included in the search parameters 131 and may be, for example, a day, a week, a month, or a period indicating a season. The component decomposition unit 113 stores the trend component, periodic component, and residual component in the component database 160. Note that the parameters used for component decomposition are included in the search parameters 131.

[0022] The component strength evaluation unit 114 calculates indices called the trend component strength and the periodic component strength for the trend component and the periodic component, respectively. The trend component strength and the periodic component strength correspond to the strength of trend and the strength of seasonality, respectively, proposed by Wang et al. (see Wang, X. et al., "Characteristic-based clustering for time series data," Data mining and knowledge Discovery, 13, pp. 335-364, 2006).

[0023] The trend component strength and periodic component strength are indices whose maximum values ​​are 1, and the closer to 1 the value is, the more desirable the component decomposition is evaluated. The more the original time series data (smoothed difference after spike removal) can be expressed with a trend component and a periodic component (the smaller the residual component), the closer the trend component strength and periodic component strength are to 1. The component intensity evaluation unit 114 stores the calculated trend component strength and periodic component strength in the component database 160, associating them with the trend component and periodic component, respectively.

[0024] The trend difference calculation unit 115 calculates the trend difference by subtracting the trend component from the power price data.

[0025] As described above, the spike prediction device 100 includes a component decomposition unit 113 that decomposes the post-spike removal smoothed differential time series data (smoothed differential after spike removal, see FIG. 5 ) into a trend component, a periodic component, and a residual component. The spike prediction device 100 also includes a trend difference calculation unit 115 that calculates trend difference data (trend difference) by subtracting the trend component from the time series data (electricity price data).

[0026] The prediction model generation unit 116 generates a spike occurrence prediction model and a spike price prediction model, which are prediction models 132. The spike occurrence prediction model is a machine learning model used to predict the time (occurrence timing) of a spike in the daily electricity price trend. The explanatory variables of the spike occurrence prediction model include a trend difference for a predetermined number of days (one or more periods of a predetermined length) before the prediction target date (prediction target period), which is the day on which a spike occurrence is to be predicted, and the occurrence date and time (occurrence timing) of the spike included in the trend difference. The occurrence date and time may be, for example, whether or not a spike occurs at each time included in the trend difference (an offset from the start of a day, which is the period, or a timing within a day). The objective variable of the spike occurrence prediction model includes the probability of a spike occurring at each time included in the prediction target date.

[0027] A spike price prediction model is a machine learning model used to predict the price at the time of a spike in daily price trends. The explanatory variables of the spike price prediction model include the trend difference for a predetermined number of days prior to the target prediction date and the date and time of the spike included in the trend difference. The date and time of the occurrence may be, for example, whether or not a spike occurs at each time included in the trend difference. The objective variable of the spike price prediction model includes the price at the time of the spike occurrence. The objective variable of the spike price prediction model may also include the price at each time on the target prediction date. This price is the price at the trend difference, and the value obtained by adding a trend component to this price is the electricity price.

[0028] The prediction model generation unit 116 generates learning data including the explanatory variables and objective variables described above by referring to the price database 140, and uses this learning data to train a machine learning model to generate a spike occurrence prediction model and a spike price model. The machine learning technology used is, for example, NGBOOST (Natural Gradient Boosting), but other machine learning technologies (regression models) such as neural networks may also be used.

[0029] The spike occurrence prediction unit 117 predicts the occurrence of a spike at each time included in the prediction target day. More specifically, the spike occurrence prediction unit 117 uses a spike occurrence prediction model based on the trend difference for a predetermined number of days before the prediction target day and the date and time of the spike occurrence included in the trend difference. If the probability is equal to or greater than a predetermined value, the spike occurrence prediction unit 117 predicts that a spike will occur.

[0030] The spike price prediction unit 118 predicts the price at the time of a spike occurrence on the target prediction date. More specifically, the spike price prediction unit 118 calculates the price at the time of spike occurrence using a spike price prediction model based on the trend difference for a predetermined number of days before the target prediction date and the date and time of the spike occurrence included in the trend difference. The spike price prediction unit 118 may also calculate the price at each time included in the target prediction date. This price is the price at the trend difference, and the value obtained by adding a trend component to this price becomes the electricity price at the time of spike occurrence.

[0031] As described above, the spike prediction device 100 includes a prediction unit (see spike occurrence prediction unit 117) that predicts the time at which a spike will occur in a prediction target period (prediction target date), which is a period (one day) for which a spike occurrence is to be predicted, based on time-series data and the time at which a spike occurred in the period prior to the prediction target period. The prediction unit predicts the time at which a spike will occur in the prediction target period based on the trend difference data (trend difference) and the time at which a spike included in the trend difference data occurred in the period prior to the prediction target period.

[0032] The prediction unit predicts the time when a spike will occur using a machine learning model (spike occurrence prediction model). The machine learning model is a machine learning model in which explanatory variables are trend difference data for a period prior to the prediction target period and whether a spike will occur in the trend difference data at each time included in the period, and the objective variable is the probability of a spike occurring at each time included in the prediction target period. The machine learning model is an NGBOOST model.

[0033] The prediction unit (see spike price prediction unit 118) predicts the value of the trend difference data at the time when a spike occurs in the prediction target period, based on the trend difference data and the time when a spike included in the trend difference data occurred in the period prior to the prediction target period.

[0034] The prediction unit predicts the value of trend difference data using a machine learning model (spike price prediction model). The machine learning model is a machine learning model in which explanatory variables are trend difference data for a period prior to the prediction target period and whether a spike will occur in the trend difference data at each time included in the prediction target period, and the objective variable is the value of the trend difference data at each time included in the prediction target period. The machine learning model is an NGBOOST model.

[0035] 6 is a flowchart of the spike extraction process according to this embodiment. At the start of the spike extraction process, candidate parameters to be used for smoothing and component decomposition are stored in the search parameters 131. Past electricity price data is also stored in the price database 140.

[0036] In step S11, the spike extraction unit 111 starts the process of repeating steps S12 to S16 for each combination of smoothing parameter candidates and component decomposition parameter candidates stored in the search parameters 131. In the following description, the smoothing parameter candidates and component decomposition parameter candidates in each iteration are referred to as smoothing parameter candidates and component decomposition parameter candidates.

[0037] In step S12, the smoothing unit 112 smooths the electricity price data to calculate a smoothing result (see FIG. 3). The parameters used for smoothing are smoothing parameter candidates. In step S13, the smoothing unit 112 subtracts the smoothing result calculated in step S12 from the electricity price data to calculate a smoothing difference (see FIG. 4).

[0038] In step S14, the spike extraction unit 111 extracts spikes included in the smoothed difference calculated in step S13 and identifies the spike occurrence times. Next, the spike extraction unit 111 removes the spikes at the spike occurrence times from the smoothed difference (see FIG. 5). In step S15, the component decomposition unit 113 performs component decomposition, which separates the smoothed difference after spike removal, into trend components, periodic components, and residual components. The parameters used for component decomposition are component decomposition parameter candidates. The component decomposition unit 113 stores the trend components, periodic components, and residual components in the component database 160 in association with the smoothing parameter candidates and component decomposition parameter candidates.

[0039] In step S16, the component intensity evaluation unit 114 calculates the trend component intensity and the periodic component intensity for the trend component and the periodic component calculated in step S15, respectively. The component intensity evaluation unit 114 stores the trend component intensity associated with the trend component and the periodic component intensity associated with the periodic component in the component database 160.

[0040] In step S17, the spike extraction unit 111 identifies the trend component for which the trend component strength, which is an index of the trend component, is maximum (best). In step S18, the trend difference calculation unit 115 calculates the trend difference by subtracting the trend component identified in step S17 from the power price data. The power price data contains spikes, and the trend difference also contains spikes. In step S19, the spike extraction unit 111 extracts spikes contained in the trend difference and stores their dates, times, and prices in the spike database 150. This price is the price in the trend difference, and the value obtained by adding the trend component to this price is the power price.

[0041] <Spike Occurrence Prediction Model Generation Process> Figure 7 is a flowchart of the spike occurrence prediction model generation process according to this embodiment. At the start of the spike occurrence prediction model generation process, the spike extraction process (see Figure 6) has been completed, and the date and time of spike occurrence has already been stored in the spike database 150. In the following explanation, the predetermined number of days related to the explanatory variables of the spike occurrence prediction model is set to 14 days. Power price data can be obtained by referring to the price database 140, and whether a spike is occurring can be obtained by referring to the spike database 150.

[0042] In step S31, the prediction model generation unit 116 starts the process of repeating steps S32 to S33 for each day. If the price database 140 contains electricity prices for four years from the 1st to the 1461st day, the prediction model generation unit 116 starts the process of repeating each of the days from the 15th to the 1461st day as a target day.

[0043] In step S32, the prediction model generation unit 116 acquires the trend difference and spike occurrence dates and times for the target day and a predetermined number of days (14 days) before that. For example, if the target day is the 1115th day, the prediction model generation unit 116 acquires the trend difference and spike occurrence dates and times for the 1115th day and for days 1101 to 1114.

[0044] In step S33, the prediction model generation unit 116 generates training data. The explanatory variables of the training data are the trend difference for a predetermined number of days before the target day and whether a spike will occur at each time included in the predetermined number of days. The objective variable is whether a spike will occur at each time included in the target day for the target variable.

[0045] In step S34, the prediction model generation unit 116 trains a machine learning model using the learning data generated in step S33 to generate a spike occurrence prediction model. The prediction model generation unit 116 generates a spike price prediction model in the same manner as the spike occurrence prediction model.

[0046] <Spike Occurrence Prediction Process> Figure 8 is a flowchart of the spike occurrence prediction process according to this embodiment. In step S41, the spike occurrence prediction unit 117 acquires the trend difference for a predetermined number of days prior to the target prediction date and the date and time of spike occurrence. In step S42, the spike occurrence prediction unit 117 predicts the spike occurrence time on the target prediction date using a spike occurrence prediction model based on the data acquired in step S41.

[0047] The spike price prediction unit 118 predicts the price using a spike price prediction model, similar to the spike occurrence prediction unit 117. Note that this price is a price at a trend difference, and the value obtained by adding a trend component to this price becomes the electricity price.

[0048] <Features of the Spike Prediction Device> The spike prediction device 100 identifies the trend component with the best trend component strength when the spike-removed smoothed difference is decomposed into components while changing the smoothing and component decomposition parameters (see steps S11 to S16 in FIG. 6 ) (see step S17). The spike prediction device 100 generates a spike occurrence prediction model and a spike price prediction model based on the spike occurrence date and time and price contained in the trend difference obtained by excluding the trend component from the electricity price data. The spike prediction device 100 predicts the spike occurrence date and price on the prediction target date using the spike occurrence prediction model and spike price prediction model. Note that this price is the price in the trend difference, and the value obtained by adding the trend component to this price becomes the electricity price.

[0049] The spike prediction device 100 predicts spikes based on a trend difference, which is a price transition obtained by subtracting a spike-removed trend component from the original electricity price. This trend component is a trend component that is not affected by spikes and indicates the original trend component of the electricity price. The spike prediction device 100 predicts spikes using a machine learning model that uses this trend difference, with the trend component removed, as training data, enabling highly accurate spike predictions that remove the influence of trends.

[0050] <<Modification: Forecast Model>> In the above-described embodiment, the forecast model 132 includes a spike occurrence forecast model and a spike price forecast model. A single forecast model 132 may be used to forecast the probability of a spike occurrence and the price. The explanatory variables of this forecast model 132 include the trend difference for a predetermined number of days (one or more periods of a predetermined length) prior to the target forecast date (target forecast period), which is the day on which a spike occurrence is to be predicted, and the date and time of the occurrence of the spike included in the trend difference. The date and time of occurrence may, for example, indicate whether or not a spike occurs at each time included in the trend difference. The objective variables include the probability of a spike occurring at each time included in the target forecast date and the price. The price is the price in the trend difference, and the value obtained by adding a trend component to this price is the electricity price.

[0051] <<Modification: Component Decomposition>> In the above-described embodiment, the component decomposition unit 113 decomposes the smoothed difference after spike removal into a trend component, a periodic component, and a residual component (see step S15 in FIG. 6 ). The power price after spike removal may also be decomposed into components. More specifically, in step S14, the spike extraction unit 111 extracts spikes included in the smoothed difference calculated in step S13, identifies the spike occurrence time, and removes the spike at the spike occurrence time from the power price data. Next, in step S15, the component decomposition unit 113 performs component decomposition to decompose the power price data after spike removal into a trend component, a periodic component, and a residual component. In this way, instead of the smoothed difference after spike removal, the power price data after spike removal may be decomposed into components to identify the trend component with the greatest trend component strength (see step S17), and the trend difference may be calculated (see step S18).

[0052] As described above, the spike extraction unit 111 extracts the time points at which spikes occurred included in the smoothed difference, which is data obtained by subtracting smoothed time series data (power price data) from the time series data, and calculates spike-removed time series data (power price data) from which the spikes at those times included in the time series data have been removed. The component decomposition unit 113 decomposes the spike-removed time series data into a trend component, a periodic component, and a residual component.

[0053] <Other Modifications> Although several embodiments of the present invention have been described above, these embodiments are merely examples and do not limit the technical scope of the present invention. In the above-described embodiments, the spike prediction device 100 predicts spikes in electricity prices, but the present invention is not limited to this and may also predict spikes in the trading prices of financial products, precious metals, grains, and other commodities. Furthermore, while the spike prediction device 100 predicts spikes using a prediction model 132 generated by itself, it may also make predictions using a prediction model 132 generated by another device.

[0054] The present invention can take on various other embodiments, and various modifications such as omissions and substitutions can be made without departing from the spirit of the present invention. These embodiments and modifications are included in the scope and spirit of the invention described in this specification, etc., and are also included in the invention described in the claims and their equivalents.

[0055] Fig. 9 is a block diagram of a computer 900 according to this embodiment. The spike prediction device 100 shown in Fig. 1 is configured by this computer 900. In Fig. 9, the computer 900 includes a CPU 901, a RAM 902, a ROM 903, an SSD 904, a communication I / F (interface) 905, an input / output I / F 906, and a media I / F 907.

[0056] The CPU 901 corresponds to the control unit 110. The RAM 902, ROM 903, and SSD 904 correspond to the storage unit 130. The communication I / F 905, input / output I / F 906, and media I / F 907 correspond to the input / output unit 180. The communication I / F 905 is connected to a communication circuit 915. The input / output I / F 906 is connected to an input / output device 916. The media I / F 907 reads and writes data from a recording medium 917.

[0057] The ROM 903 and SSD 904 store control programs executed by the CPU, various data, and the like. The CPU 901 executes application programs loaded into RAM 902 to realize various functions. The configuration of the spike prediction device 100 shown in FIG. 1 above shows the functions realized by application programs and the like as blocks. However, the spike prediction device 100 may be configured using a single computer 900, or may be configured using multiple computers 900 connected via a communication circuit 915.

[0058] REFERENCE SIGNS LIST 100 Spike prediction device 111 Spike extraction unit 112 Smoothing unit 113 Component decomposition unit 114 Component intensity evaluation unit 115 Trend difference calculation unit 116 Prediction model generation unit 117 Spike occurrence prediction unit (prediction unit) 118 Spike price prediction unit (prediction unit) 131 Search parameters 132 Prediction model 138 Program 140 Price database 150 Spike database 160 Component database

Claims

1. A spike prediction device comprising: a spike extraction unit that extracts the time at which a spike has occurred included in time series data spanning one or more periods of a predetermined length; and a prediction unit that predicts the time at which the spike will occur in a prediction target period, which is the period prior to the prediction target period in which a spike occurrence is to be predicted, based on the time series data and the time at which the spike occurred in the period.

2. The spike prediction device according to claim 1, wherein the spike extraction unit calculates post-spike-removed smoothed differential time series data, which is data obtained by removing spikes from smoothed differential data, which is data obtained by subtracting smoothed time series data from the time series data, and further comprises a component decomposition unit that decomposes the post-spike-removed smoothed differential time series data into a trend component, a periodic component, and a residual component, and a trend difference calculation unit that calculates trend difference data by subtracting the trend component from the time series data, wherein the spike extraction unit extracts a time in the period at which a spike included in the trend difference data will occur, and the prediction unit predicts the time at which the spike will occur in the prediction period based on the trend difference data and the time at which a spike included in the trend difference data occurred in the period prior to the prediction period.

3. The spike prediction device according to claim 1, further comprising: a component decomposition section that decomposes the spike-removed time series data into a trend component, a periodic component, and a residual component; and a trend difference calculation section that calculates trend difference data by subtracting the trend component from the time series data, wherein the spike extraction section extracts the time in the period at which a spike included in the trend difference data will occur, and the prediction section predicts the time at which the spike will occur in the prediction target period based on the trend difference data and the time at which a spike included in the trend difference data occurred in the period prior to the prediction target period.

4. A spike prediction device as described in claim 2 or 3, wherein the prediction unit predicts the time when the spike will occur using a machine learning model, the machine learning model being a machine learning model in which explanatory variables are the trend difference data for the period prior to the prediction target period and the likelihood of a spike occurring in the trend difference data at each time included in the period, and a dependent variable is the probability of a spike occurring at each time included in the prediction target period.

5. The spike prediction device according to claim 4, wherein the machine learning model is an NGBOOST model.

6. A spike prediction device as described in claim 2 or 3, wherein the prediction unit predicts the value of the trend difference data at the time when the spike occurs in the prediction period, based on the trend difference data and the time when a spike contained in the trend difference data occurred in the period prior to the prediction period.

7. The spike prediction device according to claim 6, wherein the prediction unit predicts the value of the trend difference data using a machine learning model, the machine learning model being a machine learning model in which explanatory variables are the trend difference data for the period prior to the prediction target period and whether a spike will occur in the trend difference data at each time included in the period, and an objective variable is the value of the trend difference data at each time included in the prediction target period.

8. The spike prediction device according to claim 7, wherein the machine learning model is an NGBOOST model.

9. A spike prediction method in which a spike prediction device executes the steps of: extracting a time at which a spike has occurred included in time series data spanning one or more periods of a predetermined length; and predicting the time at which the spike will occur in a prediction target period, which is the period for which a spike occurrence is to be predicted, based on the time series data and the time at which the spike occurred in the period prior to the prediction target period.

Images