JP2015210557A - Prediction device and prediction method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve prediction accuracy of a prediction device for predicting a future state quantity based on chaos theory.SOLUTION: A prediction device 1 predicts a future state quantity from accumulated time series data based on chaos theory. A detection part 2 detects a state quantity from a system to be controlled. A data input part 5 stores the detected state quantity in a data storage part 3 as time series data. A prediction calculation part 4 calculates a strange attractor obtained by burying the time series data stored in the data storage part 3 into a multi-dimension state space. In addition, information axes for reflecting information on factors influencing the state quantity constituting the time series data are added to the strange attractor. And then, the prediction device 1 uses various algorithms to predict the future state quantity based on the strange attractor to which the information axes have been added.

Description

  The present invention relates to a prediction technique for predicting future state quantities based on chaos theory. In particular, the present invention relates to a technology for predicting a state quantity detected in a social infrastructure or production facility.

  In social infrastructure and production facilities (plants) such as highways, power grids, and water purification plants, the system is controlled by predicting the future situation. Specifically, state quantities (for example, traffic volume, power consumption, water consumption, etc.) are acquired at regular intervals during system operation, and time-series data of the state quantities is accumulated. Future time-series data is estimated (predicted) from the time-series data of the past state quantities, and system control (for example, control of signals, ventilation fans, etc.) is performed based on the predicted future state quantities.

  In the field of predicting future state quantities based on such past state quantities and controlling the system based on the predicted state quantities, the accuracy of the prediction, the calculation speed for the prediction, and the ease of operation As a result, there will be differences in the economy, low carbonization, safety, etc. of the entire system. Therefore, the accuracy, calculation speed, simplicity, etc. of the prediction technique in system control are important factors.

  For example, in a water supply system, the amount of water demand is predicted, and a distribution plan (such as the amount of water supplied from a water treatment plant) is created according to the predicted amount of demand. Demand volume is forecasted by collecting flow rate data such as water distribution to which external conditions such as days of the week, holidays, weather, and temperature are added, and calculating the undetermined coefficient of the prediction formula based on the collected time series data. Done. In other words, the forecast calculation method predicts the demand amount for each hour unit or the demand amount for each day unit from the present, and considers the forecast demand amount and the data such as the current supply amount from the water treatment plant, and the allocation plan. Is created.

  As prediction calculation methods for predicting future demand from past demand, for example, exponential smoothing method, autoregressive method, multiple regression method, day of week exponential method, Kalman filter method and other methods based on chaos theory There are known methods (for example, Patent Documents 1-3).

  In the exponential smoothing method, normal prediction accuracy can be expected as long as there is no sudden change in the time series data. In addition, since the calculation formula is simple, data pre-processing is unnecessary.

  Since the autoregressive method uses only the past demand, it does not require any special input. Although this method can obtain a relatively high prediction accuracy, it is necessary to review the coefficients of the prediction calculation formula at regular intervals.

  The multiple regression method is a method for obtaining a predicted value by various factors. Although this method can perform prediction with high accuracy, it is necessary to review the coefficients of the prediction calculation formula at regular intervals.

  The day-of-week index method is a method for obtaining a predicted value by paying attention only to the day of the week, without taking into account various factors as in the multiple regression method. This method can be predicted with a normal degree of accuracy with a simple calculation, but it is necessary to review the day-of-week index at regular intervals.

  The Kalman filter method is used in combination with the autoregressive method or the multiple regression method, and is the most accurate method. In this method, since the prediction coefficient is automatically corrected, there is no need for correction.

  Moreover, the prediction method based on the chaos theory has high short-term prediction accuracy, and can predict future demand based on past demand.

JP-A-6-020188 JP-A-6-102939 JP-A-9-088499

  However, in an actual system, it is very difficult to keep the prediction accuracy high due to various noises such as seasonal variation, weekly variation, and irregular variation. For example, even when the Kalman filter method with high prediction accuracy is used, it is very difficult to maintain an accuracy of 90% or more.

  Further, when performing prediction based on chaos theory, prediction is performed based on determinism based on information included in time series data. This information is not separated as individual information, but is used in an unknown or unclear state. Therefore, it is difficult to use a technique such as a regression method typified by the Kalman filter method or the like to predict the combination of the two. In other words, the prediction techniques based on chaos theory and the prediction techniques such as the regression method are greatly different, so when performing a combination of the two predictions, prediction results cannot be obtained and the expected prediction accuracy is improved. It may not be recognized or the operation may be complicated.

  Moreover, the prediction based on the chaos theory has a possibility that the long-term prediction accuracy may be lower than the short-term prediction accuracy due to the characteristics of the chaos.

  In view of the above circumstances, an object of the present invention is to provide a technique that contributes to improvement of prediction accuracy in a prediction device that predicts a future state quantity based on chaos theory.

  One aspect of the prediction apparatus of the present invention that achieves the above object accumulates state quantities that change over time, and predicts future state quantities from time-series data of the accumulated state quantities based on chaos theory. An information axis for reflecting information on elements that affect the state quantity constituting the time series data is added to a strange attractor obtained by embedding the time series data in a multidimensional state space. The future state quantity is predicted based on the strange attractor to which the information axis is added.

  Also, one aspect of the prediction method of the present invention that achieves the above object is to accumulate state quantities that change over time, and to calculate future state quantities from time-series data of accumulated state quantities based on chaos theory. A prediction method for predicting, wherein a step of deriving a strange attractor in which the time series data is embedded in a multidimensional state space, and information on elements affecting the state quantity constituting the time series data in the strange attractor. A step of adding an information axis to be reflected, and a step of calculating a future data vector in which a data vector obtained by the latest observation advances based on the strange attractor to which the information axis is added using a prediction calculation method And a step of obtaining a future state quantity from the calculated future data vector.

  According to the above invention, the prediction accuracy of the prediction device that predicts the future state quantity based on the chaos theory is improved.

It is a functional block diagram of the prediction apparatus concerning the embodiment of the present invention. It is a figure which shows an example of the time series data with which it uses for the prediction of the prediction apparatus which concerns on embodiment of this invention. It is a figure which shows an example of the strange attractor produced by the prediction apparatus which concerns on embodiment of this invention. It is a figure which shows an example of the strange attractor with which possibility of selection by the prediction apparatus which concerns on embodiment of this invention increased. It is a conceptual diagram of the local fuzzy reconstruction method. It is a flowchart of the prediction method in the prediction apparatus of Example 1 of this invention. It is a figure which shows the strange attractor formed with the prediction apparatus of Example 1 of this invention. It is a figure which shows an example of the strange attractor with which possibility of selection by the prediction apparatus of Example 1 of this invention increased. It is a figure which shows an example of the strange attractor with which possibility of being selected by the prediction apparatus of Example 4 of this invention increased.

  A prediction apparatus and a prediction method according to an embodiment of the present invention will be described in detail with reference to the drawings. In addition, although description of embodiment demonstrates and demonstrates the form which performs the amount of water distribution prediction of a water purification plant, or the power consumption prediction of an electric power supply company, the prediction apparatus and prediction method of this invention are limited to embodiment. It can be applied to the prediction of state quantities that control systems in social infrastructure and production facilities such as highways and power networks (including microgrids). Further, the predicted state quantity is not limited to the state quantity for controlling the system, but the state quantity that changes over time (for example, blood glucose level, brain wave or wave height, typhoon course information, etc.) ) And can be applied to a prediction device that predicts future state quantities based on chaos theory.

  As shown in FIG. 1, the prediction device 1 according to the embodiment of the present invention includes a detection unit 2, a data storage unit 3, and a prediction calculation unit 4.

  The detection unit 2 detects state quantities in the social infrastructure and production facilities. The state quantity is an element that changes when operating a system such as a social infrastructure or a production facility, and the system is controlled according to the state quantity. For example, when performing system control of a water purification plant, the detection part 2 detects the flow volume (water distribution amount) etc. of the water used. The state quantity detected by the detection unit 2 is stored in the data storage unit 3 via the data input unit 5.

  The data storage unit 3 stores the state quantity detected by the detection unit 2 as time series data. For example, in the system of a water purification plant, the amount of water distribution that changes according to time as shown in FIG. The detected water distribution amount is accumulated for a certain period, and the accumulated amount of the water distribution amount for each predetermined period is stored in the data storage unit 3 as time series data.

The prediction calculation unit 4 predicts a future state quantity from time series data stored in the data storage unit 3 using chaos theory. More specifically, the time-series data stored in the data storage unit 3 is embedded in an n-dimensional (n is an integer of 2 or more) state space based on the Turkens embedding theorem. By this process, the data vector shown in Expression (1) is obtained. In equation (1), the embedding dimension (n) and the delay time (τ) are numbers appropriately set according to the time series data to be analyzed.
x (t) = (y (t), y (t−τ),..., y (t− (n−1) τ)) (1)
When t is changed with respect to the data vector x (t) plotted in the state space, a trajectory (strange attractor) is drawn in the n-dimensional state space. For example, a strange attractor including dynamics as shown in FIG. 3 is drawn on the n-dimensional space.

  Next, an information axis is added to the data vector x (t) plotted in the state space to reflect information on elements that affect the state quantity detected by the detection unit 2. By adding one or more information axes, the number of dimensions of the strange attractor increases according to the number of added information axes. For example, in a water purification plant system, there are factors such as year, season, day of the week, weekday, holiday, holiday, time, weather, temperature, etc. as factors that affect demand.

That is, by adding information axes (a, b, c,...) For reflecting the information of each element to the data vector x (t), a new data vector as shown in Expression (2) x ′ (t) is obtained.
x ′ (t) = (y (t), y (t−τ),... y (t− (n−1) τ), a, b, c,...) (2)
With this data vector x ′ (t), a new strange attractor is drawn in the n + m-dimensional state space (m is the number of added information axes).

  By adding an information axis in this way and drawing a new strange attractor in the state space with an increased number of dimensions, the distance between the strange attractors drawn with the data vector x (t) shown in FIG. 3 is changed. The Rukoto. As a result, as shown in FIG. 4, an attractor that is highly likely to be selected from the strange attractors drawn by the data vector x (t) is cut out. FIG. 4 shows an attractor group in which the attractor filled with a dark color has a higher probability of selection. Depending on how the information axis is taken, the strange attractor embedded in the state space with the increased number of dimensions may have a different shape from the original strange attractor. The attractor can be selected by reflecting the information of the element that gives

  Then, the data vector z (T) obtained by the latest observation is plotted on the n + m-dimensional state space. Here, the data vector x ′ (i ′) that is close to the latest data vector z (T) is referred to as a “neighbor data vector”. The neighborhood data vector x ′ (i ′) is a past observation data vector.

  The trajectory from the neighborhood data vector x ′ (i ′) of the latest data vector z (T) to the s step ahead vector x ′ (i ′ + s) is known. Assuming regularity in the fluctuation of the time series data, the trajectory from the state z (T) to z (T + s) is approximately equivalent to the trajectory from the state x ′ (i ′) to x ′ (i ′ + s). . Therefore, it is possible to select a plurality of neighboring data vectors x ′ (i ′) and calculate future data vectors z (T + s) using various algorithms (prediction calculation methods). As various algorithms, for example, Gramschmitt orthogonalization method, tessellation method, local fuzzy reconstruction method, centroid method and the like can be used. For example, in the local fuzzy reconstruction method, as shown in FIG. 5, a future data vector is predicted according to the trajectory of the neighborhood data vector. In this case, different weighting is performed depending on the distance between the latest data vector and the selected neighboring data vector. A future state quantity is derived from the predicted future data vector.

  After the future state quantity is predicted by the prediction calculation unit 4, the future state quantity derived by the prediction calculation unit 4 is stored in the prediction result storage unit 6 as shown in FIG. 1. Note that predicted values of future data vectors can be stored in the data storage unit 3 and used as teacher data.

  The state quantity predicted by the prediction calculation unit 4 is output to the output unit 7 (display or the like). The output unit 7 may display the predicted state quantity and the actually measured state quantity together. Moreover, you may display the other state quantity which fluctuates according to the estimated state quantity on the output part 7. FIG.

[Example 1]
The prediction apparatus of Example 1 predicted the demand amount of the water purification plant for a relatively long period (24 hours after the current time point). FIG. 6 shows a prediction calculation flowchart of the prediction apparatus according to the first embodiment.

  First, the water distribution amount data of the first system of the water purification plant A (data obtained by integrating the water distribution amount in the period from April 1, 2009 to March 31, 2012 every hour) is stored in the data storage unit 3. Teacher data was set (step S1).

  The teacher data stored in the data storage unit 3 is embedded in the orthogonal state space with the embedding dimension 5 and the delay time 3 (step S2). A strange attractor obtained by this embedding process is shown in FIG.

  Next, an information axis for reflecting information on elements that affect the demand amount is added to the strange attractor shown in FIG. 7 (step S3). Here, three information axes were added: seasonal information, year information, and weekday information. The number of 1 to 4 was assigned to the seasonal information axis according to each season (spring, summer, autumn and winter). Moreover, the number of 1-4 was substituted for the information axis of the year according to 2009-2012. Similarly, the numbers 1 to 3 are assigned to the weekday information axis according to weekdays, Saturdays, and Sundays and holidays. By adding the information axis, the dimension axis of the strange attractor is expanded. By predicting the future demand based on the strange attractor in which the dimension axis is expanded, as shown in FIG. 8, the strange attractor having a high probability of being selected from the strange attractor formed in Step S2 is cut out and Demand volume is predicted.

  Next, the neighborhood data vector of the latest data vector is selected on the strange attractor to which the information axis is added (step S4). Here, four data vectors are selected in the order from the latest data vector.

  From the selected four neighborhood data vectors and the trajectory of the strange attractor in the neighborhood data vector, the future data vectors of the respective neighborhood data vectors are extracted, and the local fuzzy reconstruction method, tessellation method, and Gram Schmitt orthogonalization method are extracted. The future data vector to which the latest data vector advances is extracted by the three algorithms (step S5). All algorithms were able to extract future data vectors with almost the same prediction accuracy. Regarding the prediction time, there was a time difference, and the calculation time was shorter in the order of the local fuzzy reconstruction method, the Gram Schmidt orthogonalization method, and the tessellation method. That is, when a different algorithm is used, although the prediction calculation time is different, a future data vector can be predicted with substantially the same prediction accuracy (the same is true for the other embodiments).

  Finally, the predicted value of the future demand is calculated from the future data vector extracted in step S5 (step S6).

  As a control experiment, the demand amount was predicted by the same method as the prediction apparatus of Example 1 except that the addition of the information axis (step S3) was not performed.

For the prediction of the demand amount, the total demand amount from the current time to 24 hours ahead was predicted. For the prediction up to 24 hours ahead, the demand amount is predicted every hour, and the obtained demand amount is integrated for 24 hours to obtain the demand amount up to 24 hours ahead. The demand amount was predicted from April 1, 2012 to March 31, 2013, and the prediction accuracy of the demand amount was evaluated based on Equation (3). Generally, in a water purification plant, water distribution is performed according to the demand amount, and therefore, the “predicted water distribution amount” in the equation (3) is the same amount as the “predicted demand amount” (the same applies to other embodiments). The demand amount was predicted for one year every day, and the absolute value of the prediction accuracy obtained in one year was averaged.
Prediction accuracy = ((Actual water distribution-Predicted water distribution) / Actual water distribution) x 100 (%) (3)
The prediction accuracy in the prediction device of the control experiment was 1.93. On the other hand, the prediction accuracy in the prediction apparatus of Example 1 is 1.84, and the prediction accuracy is improved by adding the information axis.

[Example 2]
The prediction device of the second embodiment predicts the demand amount in the same manner as the prediction device of the first embodiment, except that the demand amount is predicted for a relatively short period (one hour after the current time). It was. Similar to the first embodiment, the prediction apparatus according to the second embodiment predicts demand according to the flowchart illustrated in FIG. 6.

  First, the water distribution amount data of the second system of the water purification plant A (data obtained by integrating the water distribution amount from April 1, 2009 to March 31, 2012 every hour) is stored in the data storage unit 3, and the teacher data (Step S1).

  The teacher data stored in the data storage unit 3 is embedded in the orthogonal state space with the embedding dimension 5 and the delay time 3 (step S2).

  Three information axes of seasonal information, year information, and weekday information were added to the embedded strange attractor (step S3).

  Four neighboring data vectors of the latest data vector are selected on the strange attractor to which the information axis is added (step S4), and the future demand is predicted based on the selected neighboring data vector (step S5). , S6).

  As a control experiment, the demand amount was predicted by the same method as the prediction apparatus of Example 2 except that the addition of the information axis (step S3) was not performed.

  The prediction of the demand amount predicted the total demand amount from the current time to one hour ahead. The demand amount was predicted from April 1, 2012 to March 31, 2013, and the prediction accuracy of the demand amount was evaluated based on Equation (3). The demand amount was predicted every hour for one year, and the absolute value of the prediction accuracy obtained in one year was averaged.

  The prediction accuracy in the prediction device of the control experiment was 5.76. On the other hand, the prediction accuracy in the prediction apparatus of Example 2 was 5.14, and the prediction accuracy was improved by adding the information axis.

[Example 3]
The prediction device of the third embodiment predicts the demand amount of the water purification plant by the same method as the prediction device of the first embodiment, except that the elements of the information axis added to the strange attractor in which the time series data is embedded are different. It is a thing. The prediction device of Example 3 predicted the demand amount according to the flowchart shown in FIG.

  First, the water distribution amount data of the first system of the water purification plant B (data obtained by integrating the water distribution amount from April 1, 2010 to March 31, 2012 every hour) is stored in the data storage unit 3, and teacher data (Step S1).

  The teacher data stored in the data storage unit 3 is embedded in the orthogonal state space with the embedding dimension 5 and the delay time 3 (step S2).

  Three information axes were added to the embedded strange attractor: time information, temperature information, and day information. A number of 1 to 24 was assigned to the time information axis according to the time. Moreover, the number of 1-3 was substituted for the low temperature, the intermediate temperature, and the high temperature on the information axis of the air temperature. Similarly, the numbers 1 to 7 are assigned to the information axis of the day of the week according to Monday to Sunday.

  On the strange attractor to which the information axis is added, the five nearest neighbor data vectors are selected from the latest data vector (step S4), and the future demand is predicted based on the selected neighbor data vector ( Steps S5 and S6).

  As a control experiment, the demand amount was predicted by the same method as in Example 3 except that the addition of the information axis (step S3) was not performed.

  For the prediction of the demand amount, the total demand amount from the current time to 24 hours ahead was predicted. For the prediction up to 24 hours ahead, the demand amount is predicted every hour, and the obtained demand amount is integrated for 24 hours to obtain the demand amount up to 24 hours ahead. The demand amount was predicted from April 1, 2012 to March 31, 2013, and the prediction accuracy of the demand amount was evaluated based on Equation (3). The demand amount was predicted every 24 hours for one year, and the absolute value of the prediction accuracy obtained in one year was averaged.

  The prediction accuracy in the prediction device of the control experiment was 3.46%. On the other hand, the prediction accuracy in the prediction apparatus of Example 3 was 3.18%, and the prediction accuracy was improved by adding an information axis.

[Example 4]
The prediction apparatus according to the fourth embodiment predicts the power consumption of the power supply system. The prediction apparatus according to the fourth embodiment is different from the prediction apparatus according to the first embodiment in data stored in the data storage unit 3 in step S1. In the prediction apparatus according to the fourth embodiment, power consumption is predicted according to the flowchart shown in FIG.

  First, power consumption data of the power supply system (data obtained by accumulating the power consumption from April 1, 2008 to March 31, 2012 every hour) is stored in the data storage unit 3 and used as teacher data. (Corresponding to step S1).

  The teacher data stored in the data storage unit 3 is embedded in the orthogonal system state space with the embedding dimension 5 and the delay time 4 (step S2).

  Three information axes of weather information, temperature information, and weekday information were added to the embedded strange attractor (step S3). Numbers 1 to 3 were assigned to the weather information axis according to sunny, cloudy, and rainy (snow). Moreover, the number of 1-3 was substituted for the low temperature, the intermediate temperature, and the high temperature on the information axis of the air temperature. Similarly, the numbers 1 to 3 are assigned to the weekday information axis according to weekdays, Saturdays, and Sundays and holidays. FIG. 9 shows an example of an attractor that is more likely to be selected by adding an information axis.

  On the strange attractor to which the information axis is added, the nearest three neighboring data vectors are selected from the latest data vector (step S4), and the future power consumption is predicted based on the selected neighboring data vector. (Steps S5 and S6).

  As a control experiment, power consumption was predicted by the same method as the prediction apparatus of Example 4 except that the addition of the information axis (step S3) was not performed.

For the prediction of power consumption, the total power consumption from the current time to 24 hours ahead was predicted. For prediction up to 24 hours ahead, power consumption is predicted every hour, and the obtained power consumption is integrated for 24 hours to obtain power consumption up to 24 hours ahead. The prediction of power consumption was performed from April 1, 2012 to March 31, 2013, and the evaluation of the prediction accuracy of the power consumption was performed based on Expression (4). The power consumption was predicted every 24 hours for one year, and the prediction accuracy was evaluated by averaging the absolute values of the prediction accuracy obtained in one year.
Prediction accuracy = ((actual power consumption−predicted power consumption) / actual power consumption) × 100 (%) (4)
The prediction accuracy in the control device of the control experiment was 3.61%. On the other hand, the prediction accuracy in the prediction apparatus of Example 4 was 3.13%, and the prediction accuracy was improved by adding the information axis.

  According to the prediction device and the prediction method according to the embodiment of the present invention as described above, the prediction accuracy of the prediction device that predicts the future state quantity based on the chaos theory is improved. That is, it is possible to predict the future state quantity (data vector) by reflecting the element information so as to make a more optimal selection while keeping the goodness of the algorithm of selecting an appropriate data vector from the strange attractor.

  According to the prediction device and the prediction method according to the embodiment of the present invention, the strange attractor obtained based on the chaos theory and the various axes related to the state quantities constituting the strange attractor are integrated into a dimension axis. By doing so, it is possible to draw a hybrid attractor that makes use of both the function of regression prediction from known information and information based on chaos theory (prediction based on determinism) developed in the strange attractor. As a result, it is possible to predict the future state quantity by adjusting the contribution of various information related to the state quantity from the obtained attractor, improving the prediction accuracy in a wide range from short-term prediction to long-term prediction Can be made.

  In particular, in the prediction apparatus and the prediction method according to the embodiment of the present invention, prediction accuracy can be improved not only in short-term prediction but also in long-term prediction. As a result, it is possible to improve the accuracy of long-term prediction in prediction based on chaos theory, which generally has higher accuracy in short-term prediction than in long-term prediction.

  The description of the embodiments of the present invention has been given by way of specific preferred examples. However, the present invention is not limited to the examples, and design changes may be made as appropriate without departing from the characteristics of the invention. Possible and modified forms are also within the technical scope of the present invention.

  For example, the prediction device of the present invention is applied to prediction of state quantities for controlling each system in a social infrastructure such as an expressway and a power network (including a microgrid) and production facilities. Therefore, the prediction apparatus of this invention can also be used as a control apparatus which controls each system based on the predicted state quantity.

  In the description of the embodiment, the future state quantity is predicted based on the time series data of the past state quantity. However, the future state quantity is calculated in real time from the time series data of the past state quantity and the current state quantity. The amount can also be determined.

  In addition, elements added as information axes are not limited to those in the embodiment, and elements that can be measured such as temperature and those that can be determined such as weather can be used. Specifically, year (2014, etc.), month (January to December), day (1st to 31st), time or time (1 hour to 24 hours), minute (1 minute to 60 minutes), Information on calendar and time such as seconds (1 to 60 seconds), atmospheric pressure (hPa), precipitation (mm), temperature (° C), humidity (%), wind direction / wind speed (m / s), sunshine duration (h ), Snowfall (cm), snowfall (cm), weather (clear, clear, cloudy, rain, snow, sandstorm, etc.), solar radiation (direct solar radiation, diffuse solar radiation, global solar radiation) Information related to weather and astronomical phenomena such as tide level, UV, and pollen, yellow sand, PM2.5 concentration of fine particulate matter, weekdays, etc. (holidays, holidays), seasons, etc. It is possible to add.

DESCRIPTION OF SYMBOLS 1 ... Prediction apparatus 2 ... Detection part 3 ... Data storage part 4 ... Prediction calculation part 5 ... Data input part 6 ... Prediction result storage part 7 ... Output part

Claims (4)

Accumulate state quantities that change over time,
A prediction device that predicts future state quantities from accumulated time series data based on chaos theory,
In the strange attractor obtained by embedding the time series data in a multidimensional state space, an information axis for reflecting information on elements that influence the state quantity constituting the time series data is added,
A prediction apparatus for predicting the future state quantity based on a strange attractor to which the information axis is added. Accumulate state quantities of the system to be controlled at regular intervals,
A prediction device that predicts future state quantities from accumulated time series data based on chaos theory,
In the strange attractor obtained by embedding the time series data in a multidimensional state space, an information axis for reflecting information on elements that influence the state quantity constituting the time series data is added,
A prediction apparatus for predicting the future state quantity based on a strange attractor to which the information axis is added. A plurality of neighboring data vectors that are close to the data vector obtained by the latest observation are selected from the strange attractor to which the information axis is added, and the latest observation is performed using a prediction calculation method from the selected neighboring data vector. Calculate the future data vector that the data vector obtained in step
The prediction apparatus according to claim 1, wherein the future state quantity is derived from the future data vector. A method of accumulating state quantities that change over time, and predicting future state quantities from time-series data of accumulated state quantities based on chaos theory,
Deriving a strange attractor in which the time series data is embedded in a multidimensional state space;
Adding to the strange attractor an information axis for reflecting information of an element that affects the state quantity constituting the time-series data; and
Based on the strange attractor to which the information axis is added, calculating a future data vector in which the data vector obtained by the latest observation advances using a prediction calculation method;
A step of obtaining a future state quantity from the calculated future data vector.