JP2003108909A - Short-term prediction system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a short-term prediction system predicting a change of a prediction term on the basis of a plurality of observation results of an observation term in the past as to a continuously varying phenomenon in a short term. SOLUTION: An m+1 unit time before the prediction term to be predicted in the future is used as the observation term. Among the observation results in the past, observation results of n pieces of observation terms having similar characteristics with the observation term are represented by Q1-Qn. Observation results of the observation terms after a M time unit from these observation results Q1 to Qn are represented by R1 to Rn. Assuming that a change during a term 0<t<=0 as the prediction term of X(t) is decided by a change during a term X(t) duration - T0 <=t<=0 if T1 is comparatively small, an observation result having similar characteristics to a change in the term X(t) duration -T0 <=t<=0 is extracted from the observation results in the past, and the n pieces of the observation results Q1 to Qn are obtained. From the extracted observation results, a change in the prediction term is predicted by computing.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a short-term prediction method for continuously changing phenomena.

[0002]

[Technical background] 1) Prediction technology using mathematical methods Natural phenomena are phenomena that have periodicity, such as monthly changes in weather that depend on the Earth's revolution, and tides that depend on the Moon's revolution.
It is broadly divided into non-periodic phenomena such as weather changes for several days. Among these, as the mathematical theory that can be applied to the efficient solution of the problem of the periodic phenomenon, the theory of time series and the theory of Markov process can be considered. However, for the phenomenon without periodicity, there is currently no theory of application method for solving this kind of problem. 2) Short-term forecasting technology Numerical forecasting is a calculation of the time-based change in meteorological quantities for each three-dimensional grid point in the forecasted area, using fluid dynamics, thermodynamics, and other laws of physics. This calculation is based on a numerical forecast called an area model, and is created using the relational expression between the statistically obtained numerical calculation results and the actual weather and temperature. Commonly used numerical forecasts are MOS (Model Output Statistics) and PPM (Perfec
t Prog. Methods). The former is the statistical value between the value of the weather element predicted by the numerical forecast and the value of the actual observed weather element or the index used to represent the weather that cannot be expressed by the weather element such as sunny or cloudy. Create a relational expression,
It is a method of substituting the daily expected value into this relational expression and predicting the value of the final meteorological element and the index used to represent the weather. In the latter, a statistical relational expression between the actual value (or the initial value of the numerical forecast) and the value of the corresponding meteorological element or the index used to express the weather is prepared, and the daily expected value is calculated using this relationship. This is a method of predicting the value of the final meteorological element and the index used to represent the weather by substituting it into the formula. In the case of MOS, there is an advantage that the systematic error of the numerical forecast is automatically corrected, and the accuracy is higher than that of PPM.

[0003]

As the statistical relational expression used in the above-mentioned conventional method, there is a multiple correlation regression expression in which an expected value of a numerical forecast is used as an explanatory variable. In that case, it is necessary to accumulate numerical forecasts for at least the past two to three years, and it is not possible to immediately respond to upgrading the numerical forecast model. In addition, Kalman filter and neural network are used as the method. These methods are
It is equipped with a learning function that automatically corrects the statistical relational expression by comparing the actual weather with the index used to represent the weather obtained based on the numerical forecast. Therefore, a new model can be dealt with within a few months, but if the atmospheric field fluctuates significantly from the situation during the learned period, the accuracy may temporarily decrease until the new situation is sufficiently learned. Further, there are many problems in improving the accuracy of the forecast by the above-mentioned method, such as improving the accuracy of numerical forecast output and improving the accuracy of translation into statistical weather such as MOS. The present invention has been made in view of the above points, devises a new time series estimation method, solves the above conventional problems, and in the case of weather, for example, the temperature, wind direction up to 72 hours later, Wind speed, barometric pressure,
It is an object of the present invention to provide a method for simultaneously predicting changes in humidity and the like.

[0004]

In order to achieve the above-mentioned object, the present invention provides a method for measuring a phenomenon that continuously changes for a short period of time based on observation results of a plurality of past observation periods, It is a short-term prediction system that performs prediction, and from the past observation results, the observation result Q of the observation period (-mt ₀ <t _i ≤ 0: t ₀ is a unit time) and the observation result R of the period corresponding to the prediction period A storage means for storing a set of observation results as S, and an observation result for each unit time t _{0 of} a plurality (n) of S as a matrix X (t _i ), and for each S of the stored observation results. Average of the corresponding row vectors X _α (t _i ) per unit time

[Equation 9] To calculate the relationship between the calculated mean value and the difference between each observation result

[Equation 10] And W for the time t _{k in the} prediction period.
simultaneous equations that correlate _{i and j}

[Equation 11] Solve, respective coefficients _{A 0,} A 1, · · ·, means for determining the _{A m,} using a solution of the simultaneous equations, the value Y _{(t k)} at time _{t k} of the prediction period

[Equation 12] In the short-term prediction system, the obtained y (t _k ) is output as a prediction value in the prediction period. The observation result can be at least one state of meteorology, sea state, and geography observed at a plurality of observation points where observation is constantly performed.
The present invention also includes a program that enables a computer system to implement the short-term prediction system described above, a recording medium that stores the program, and a method performed by the short-term prediction system described above.

[0005]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described in detail below with reference to the drawings. Throughout this specification, the vector with the superscript T represents a transposed vector.
The short-term prediction method according to the present invention is executed by using a general computer system. FIG. 1 is a block diagram showing a configuration example of hardware for implementing the short-term prediction method of the present invention. In FIG. 1, the processing device main body 20 includes a CPU 22, a memory 24, each interface 28, and an auxiliary storage device 26 such as a disk connected to one of the interfaces 28. In addition, an input device (keyboard or the like) 3 via the interface 28
0, connected to an output device (display or the like) 40. The system of the present invention is stored in the auxiliary storage device 26 as a program that realizes the functions of the present invention, and the input device 3
When activated by an instruction from 0, it is loaded on the memory 24 and executed by the CPU 22. The execution result is output to the output device 40 such as a display or a printer.

FIG. 2 is a diagram for explaining extraction of past observation results for use in the short-term prediction method of the present invention. An embodiment of the present invention will be described using weather information as an example.
In FIG. 2, on the time axis shown in FIG.
February 4th. Now, let m + 1 unit time before the prediction period to be predicted in the future be the observation period, and in the past observation results shown in FIG. 2B, the past n with similar characteristics to the observation period. The observation result for each observation period is Q ₁ ,
_{Q 2, ···, Q n,} R 1 observation result of the observation period of M unit time after _{them, R 2, ···, R n} , S 1 observation results summarized _{these, S 2,} · .., S _n , the number of observation points is d, and the natural phenomenon to be observed is p species. At this time,
A vector X representing the state at time t for each S
(T) is a pd-dimensional vector

[Equation 13] It is represented by. However, i = -m, ..., -1, 0, 1,
2, ..., M. This represents the state of the weather of interest at each point at time t at the same time.
For example, when considering the temperature and humidity in Kanazawa, Tokyo, and Sendai, the number of observation points is d = 3, and the natural phenomenon type p = 2. As illustrated in FIG. 2 as information on December 1st and December 4th, (x _1,1 (t), x _1,2 (t)) is the temperature and humidity of Kanazawa at time t, and (x _{2 , 1} (t), x
_{2, 2} (t)) is the temperature and humidity of Tokyo at time t,
(X _3,1 (t), x _3,2 (t)) represent the temperature and humidity of Sendai at time t. X (t) is a 6-dimensional row vector that comprehensively represents these.

The change in the prediction period of X (t), 0 <t ≦ T ₁ , is determined by the change in the period of X (t) −T ₀ ≦ t ≦ 0 when T ₁ is relatively small. Is assumed to be
Observation results having characteristics similar to those in the period (t) −T ₀ ≦ t ≦ 0 are extracted from past observation results, and _n observation results Q ₁ , Q ₂ , ..., Q n are obtained. To do. Regarding the selection of past observation results having similar characteristics, the similarity does not have to be strict, and the observation results in which the transitions of the observed values are slightly similar or the observation results at the same time may be used. The extraction may be performed by detecting the similarity by pattern matching, or may be performed simply by specifying the same period as the observation period.

FIG. 3 is a diagram for explaining the arrangement of the extracted observation results. In FIG. 3A, first, in the extracted observation results Q ₁ , Q ₂ , ..., Q _n , each period M
Observations _{R 1,} R 2 corresponding to 0 <t ≦ _{T 1} is the prediction period after, ..., and _{R n} In addition, collectively _S 1,
Let S ₂ , ..., S _n . Then, the observation period t _i = it ₀ (i = −m,
..., -1, 0, 1, 2, ..., M),
Because observations X (t _i) the combined matrix of the past observation period (see FIG. 3 (b)) is obtained, the prediction is performed using the same. However, in FIG. 3, t ₀ is a unit time, and t _−m = −T ₀ and t _M = T ₁ .

FIG. 4 is a flow chart for explaining a process for making a prediction using the arranged observation results.
To make the prediction, the above-mentioned −T ₀ ≦ t ≦ T _{1 of} X (t).
The "habit" of the change in the observation results during the period is investigated and the prediction is made based on that. The processing procedure will be described in detail below with reference to the flowchart of FIG. In FIG. 4, first, as described above, the observation results S ₁ , S ₂ , ...
., P corresponding to S _n , observed at point d at time t _i
A row vector X (t _i ) that summarizes the observation results of the seeds is obtained (S102). Next, from the row vector X (t _i ), a bar vector X (t _i ) which is a row vector whose element is the average of each element of the n cases is obtained.

[Equation 14] (S104). However, X _α (t _i ) is the time t _i (i = _i) in the α-th case out of n cases extracted from past observation cases and having similar characteristics to the observation result in the observation period before prediction. -M, ...,-1,0,1,
2, ..., M) are vectors representing the states.

To obtain the relationship between the difference between the obtained average value and the observed case, bar X (t _i ) which is the average value of the observation result X (t _i ) at time t _i for −m ≦ i, j ≦ M. ) And the transposed matrix of the matrix of the difference is X (t _j ) at time t _j
The value W _{i, j} for each case of the value obtained by multiplying the matrix of the difference with the bar X (t _j ) that is the average value of

[Equation 15] (S106). However, X _α (t _i ) is the time t _i (i = _i) in the α-th case out of n cases extracted from past observation cases and having similar characteristics to the observation result in the observation period before prediction. -M, ...,-1,0,1,
2, ..., M) are vectors representing the states.

The obtained W _{i, j} is the observation result X (t _i ).
Average value in the form of the relationship of the difference between the bar X (t _i) the difference between the average value of X (t _j) of that bar X (t _j), that observations changes in expressed by the difference between the average value Represents the "habit" of. Therefore, the difference between the average value of the past observation results R corresponding to the time t _k of the prediction period and the times t _−m and t _{−m of the} previous observation period.
, Relation of difference with the average value of the observation result Q at _{+ 0} , ..., T ₀ W _{k, −m} , W _{k , −m + 1} , ..., W _{k, 0}
Is a system of equations that correlate W _{i, j}

[Equation 16] Can be expressed as However, W _{i, j} is a value calculated by the above equation (3). In addition, _{t k} is the k-th time of the forecast period (k = 1,2, ···, M ) is.

[0012] by solving the simultaneous equations, the coefficients of the linear combination _{A 0, A 1, ···,} determine the _{A m} (S108). The obtained A ₀ , A ₁ , ..., _Am are the time t _{k of the} prediction period.
, The time t ₀ , t ₋₁ , ..., Of the previous observation period.
It is an estimated value of the degree of influence of the difference from the average value of the observation results at t− _m . Therefore, when this is multiplied by the observation result of the prediction period and the observation period before that, the prediction result of the difference from the average value at the time t _k of the prediction period is obtained, and the final prediction result is obtained by adding the average value to it. can get. Finally, every time t _l (−m ≦ l ≦ 0) of the observation period before the prediction period,
The solutions A _l of simultaneous equations calculates the sum over the transposed matrix of a matrix of the differences between the X bar X (t _l) of (t _l), row vector bar represents the average of past observations corresponding thereto X
A transposed matrix of _{(t k)} of the transposed matrix were added resulting matrix Y _{(t k)}

[Equation 17] (S110). However, m + 1 mpd × m
pd matrix A ₀ (t _k ), A ₁ (t _k ), ..., A
_m (t _k ) is a value calculated by the above equation (4). In addition, _{t k} is the k-th time of the forecast period (k = 1,
2, ..., M). The element of the matrix Y (t _k ) obtained by the above calculation is the predicted value at the time t _k of the prediction period.

Further, when p = 1 and d = 1, each time t
error of the observed and predicted values of _k are mean 0, variance sigma _k ²
Assuming that each error is independent of each other and the error is independent, the confidence band of the confidence coefficient 1-α (0 <α <1) is two polygonal lines A and B shown in FIG.

[Equation 18] It becomes the area surrounded by. However, l is called an α point, and in the case of normal distribution, 1.9 is obtained when 1-α is 95%.
It is 6. Note that FIG. 5 shows a case where p = 1 and d = 1.

By setting the observation period and the prediction period, it is possible to omit the calculation and make a longer-term prediction. For example, when the observation period and the prediction period are separated from each other as shown in FIG. 6 (two days are left apart in FIG. 6), the prediction immediately after the observation period does not have to be performed, so that the prediction with a smaller calculation amount can be performed earlier. It is possible to do. In this case, the shorter the distance between the observation period and the prediction period, the higher the prediction accuracy, but this prediction method can be applied as long as the influence of the observation period extends to the prediction period.

Further, as shown in FIG. 7, it is possible to carry out further prediction by sequentially incorporating the prediction results into the observation period and repeating the procedure shown in the flowchart of FIG. In FIG. 7, the prediction period is Mt ₀ and the observation period is mt ₀ . FIG. 7 shows a case where M = 3, m = 3, and unit time t ₀ = 1 day. In FIG. 7, first, as a prediction, the observation period (December 1
It shows that the prediction for the prediction period (December 5 to December 7) is performed from the date to December 4). Next, when predicting the future prediction period (from December 8 to December 10), the observation period is from December 4 to December 7, but the observation result was previously predicted. The prediction is performed by incorporating the prediction results of the prediction period (December 5th to December 7th) (prediction). Also in the prediction, the predicted result is similarly used to perform the prediction from December 11 to December 13, which is the prediction period. In this case, set a long observation period and prediction period in advance, and
₀ (t _k ), ..., A _m (t _k ) are obtained by the expression (4), and the bar X (t _k ) and X (t ₀ ) −bar X (t _l ) of the expression (5) are _obtained. The calculation can be omitted by replacing (-m ≦ l ≦ 0) with a new value.

Further, as shown in FIG. 8, it is possible to make a further prediction by sequentially adding prediction results to the observation period and repeating the processing procedure shown in FIG. In FIG. 8, the prediction period is Mt ₀ , the observation period is m ₁ t ₀ ,
m ₂ t ₀ and m ₃ t ₀ . The unit period t ₀ is one day. As shown in FIG. 8, first, by performing the above-described process in the prediction, the prediction period (from December 5 to 12 in FIG.
Forecast (until the 7th of the month). Next, as shown in FIG. 8B, the prediction of the obtained prediction period is incorporated as the observation result of the observation period, and the prediction period after that (from December 8 to December 10 in FIG. 8) , And the same prediction is performed again (prediction). Thereby, the prediction result of the prediction can be reflected in the prediction result of the prediction. Similarly, as shown in FIG. 8C, the prediction can be reflected and the prediction for the subsequent prediction period (from December 11 to December 13) can be performed. In this case, a long observation period and prediction period are set in advance, and A ₀ (t _k ), ..., _Am
(T _k ) is obtained by the equation (4), and the bar X (t _k ) is obtained.
(See formula (5)) and X (t ₀ ) -bar X (t ₁ ) (-m
By replacing ≦ l ≦ 0) with a new value, the calculation can be omitted.

[0017]

EXAMPLE An example in which the above processing is applied to a specific event will be described below. The temperature, wind direction, wind speed, and humidity from 2:00 on December 29 to 0:00 on January 1 are 15 such as Kanazawa, Tokyo, Sendai, etc.
Using the result of observation at the point, from 3:00 on January 1 to January 3
For example, a case of predicting every 3 hours until 24:00 will be taken as an example. Therefore, the data from 72 hours before is used to predict up to 72 hours after every 3 hours. In this case, there are four types of weather information p and 15 points d. First, number the 15 points. In addition, numbers are attached to the weather to be predicted, that is, temperature, wind direction, wind speed, and humidity. The unit time is 3 hours. According to the above symbols, t _i = 3i, (i = −23, ..., 0, ...,
24). Therefore, the observation time is t ₋₂₃ = −6
9 (hours), ..., t ₀ = 0 (hours), and the predicted time is represented as t ₁ = 3 (hours), ..., t ₂₄ = 72 (hours). Here, one year is divided as follows. January 1st-January 15th January 16th-January 31st February 1st-February 14th February 15th-February 28th (or 29th) ... December 16th-12th 31st of the month This corresponds to S in FIG.

Next, in order to perform the prediction in S, the data is arranged as described in FIG. The data shall be obtained every 10 minutes. The average values of temperature, wind direction, wind speed, and humidity obtained around 30 minutes on December 29 of the previous year at 3:00 are calculated for all the observation points. Therefore, time t
_{At −23,} the state at the j-th point (j = 1, ..., 15) is (x _{j, 1} (t ₋₂₃ ), x
_{j, 2} (t- ₂₃ ), _{xj , 3} (t- ₂₃ ), _{xj, 4}
(T- ₂₃ )). In addition, the vector X (t- ₂₃ ) of all the observed values at time t- ₂₃ is

[Formula 19] Can be expressed as Consider the same at other times.
Then, the mean bar X (t _i ) and the “covariance” W
_{i and j} are obtained by the above equations (2) and (3).
Here, since there is data for the past 20 years, the data is organized as follows. The above method is used to calculate the data for each of the past 20 years. That is, the α year (= 1, ...
.., 20) and X _α (t) is calculated by the same calculation as above.
Ask for. X _α (t ₋₂₃ ) is calculated from the data at _3:00 on December 29. X _α (t ₋₂₂ ) is calculated from the data at 6:00 on December 29. ... X _α (t ₀ ) is calculated from the data at 0:00 on January 1st. X _α (t ₁ ) is calculated from the data at 3:00 on _{January 1st} .・ ・ ・ X _α (t ₂₄ ) is calculated from the data at 24:00 on January 3rd.

After that, X _α (t− ₂₃ ), ..., X _α (t ₀ ), ... Corresponding to the data for 20 years obtained above.
.., each t of X _α (t ₂₄ ) (α = 1, ..., 20)
Average value at _k (k = −23, ..., 0, ..., 24)

[Equation 20] To calculate. However, α is the α year (= 1, 2, ...,
20) is shown. This is a vector that comprehensively shows the average values of the temperature, wind direction, wind speed, and humidity at each observation point.

To determine W _{i, j} ,

[Equation 21] To calculate. Then, to obtain the "mean" bar X (t _k ), the "covariance" W _{i, j} , using the method described above, the bar X (t _k ), W _{i, j} corresponding to each of the following periods _: To calculate.・ The bar X (t _k ), W _{i, j} corresponding to 3:00 on December 29th to 24:00 on January 3rd in the previous year is obtained, and the bar X ⁽¹⁾ (t
_k ), W _{i, j} ⁽¹⁾ . -Determine bar X (t _k ), W _{i, j} corresponding to 3:00 on December 30th to 24:00 on January 4th, _{and calculate} bar X ⁽²⁾ (t
_k ), W _{i, j} ⁽²⁾ . -Determine bar X (t _k ), W _{i, j} corresponding to 3 o'clock on December 31st to 24:00 on January 5th, and obtain bar X ⁽³⁾ (t
_k ), W _{i, j} ⁽³⁾ . ..... bar X (t _k ), W _{i, j} corresponding to 3:00 on January 10th to 24:00 on January 15th, is obtained.
Put the ^{_{(15) (t k),}} W i, j (15).

Next, the bar X obtained above
^{(1) using} (t _k ), W _{i, j} ^(1), etc.

[Equation 22] And calculate this as the bar X required in equation (3).
(T _k ) and W _{i, j} . In the above formula, the bar X (t _k ) and W _{i, j} are obtained by shifting the moving average by one day and stabilizing the respective values.

With m = 23 and M = 24, for each k (k = 1, 2, ..., 24) by the above method, the following 24
24 matrices A ₀ (t _k ), A ₁ (t _k ), which are determined as solutions of simultaneous equations (see equation (4))
_..., determine the A 23 _{(t k).} In order to perform the prediction until 72 hours later, the row vectors representing the 24 observation values before the prediction is started are X (t- ₂₃ ), X (t- ₂₂ ), ...
, X (t ₋₁ ), X (t ₀ ), and the 24 matrices A ₀ (t _k ), A ₁ (t _k ), ..., A ₂₃ (t
_k ) is used to calculate the transposed matrix of the matrix Y (t _k ) according to equation (5). Finally, time points t ₁ , t ₂ , ..., T ₂₄
, Y (t ₁ ), Y (t ₂ ), ..., Y in
Calculate (t ₂₄ ). However,

[Equation 23] Is.

When using the above results to predict until 72 hours later, for example, in Tokyo, the following prediction is made.・ Temperature, wind direction, wind speed, humidity at 3:00 on January 1st: y
_2,1 (t ₁ ), y _2,2 (t ₁ ), y
_2,3 (t ₁ ), y _2,4 (t ₁ ) -Temperature, wind direction, wind speed, humidity at 6:00 on January 1st: y
_2,1 (t ₂ ), y _2,2 (t ₂ ), y
_2,3 (t ₂ ), y _2,4 (t ₂ ) ... Temperature, wind direction, wind speed and humidity at 24:00 on January 3rd: y _2,1
(T ₂₄ ), y _2,2 (t ₂₄ ), y
_2,3 (t ₂₄ ), y _2,4 (t ₂₄ ) In this case, M = 24. Therefore, if l = 3.1,
Even considering the estimation of each σ _i , the reliability coefficient is about 70%.
Can get the confidence band of. January 1st 3:00 to January 1st
The prediction up to 24:00 on the 5th is based on A _i (t _k ) and W obtained above.
Calculate Y (t _i ) using _{i, j} (t _k ). For example, when using data from 2:00 on January 2 to 0:00 on January 5 to predict from 3:00 on January 5 to 2:00 on January 8, t- _{23 is set} to 2:00 on January 2 You can think of it in correspondence.

9 to 11 show the above-mentioned procedure in Kanazawa,
A table showing the predicted values and the observed values created based on the results of predicting the temperature of the next day from December 4, 1990 to December 15, 1990 using the observation results every three hours at three points in Tokyo and Sendai. It is a graph. In the results shown in these figures, Kanazawa,
The square root (RMSE) of the mean squared error (RMSE) between the predicted values and the observation points that considered Tokyo and Sendai at the same time was 0.24, and the error exceeded 1 degree (1.02 degree) was predicted 288 times. It shows that it is only once in.

[0025]

According to the present invention, for example, based on the observation results at a plurality of points, it is possible to predict a natural phenomenon occurring after a short period together with a confidence interval.
There is an effect that a more accurate and practical prediction can be obtained with a smaller calculation amount. The present invention is useful as a countermeasure for each industry affected by such a natural phenomenon.

[Brief description of drawings]

FIG. 1 is a system configuration diagram in which the present system is implemented.

FIG. 2 is a diagram illustrating a prediction period, an observation period, and the like.

FIG. 3 is a diagram for explaining the arrangement of past observation results.

FIG. 4 is a flowchart illustrating a process of the embodiment.

FIG. 5 is a diagram showing the reliability of the prediction result of this system.

FIG. 6 is a diagram illustrating another method of taking a prediction period and an observation period.

FIG. 7 is a diagram illustrating that prediction is incorporated and further prediction is performed.

FIG. 8 is a diagram illustrating another case in which prediction is incorporated to make further prediction.

FIG. 9 is a table showing results of prediction using actual observation results.

FIG. 10 is a table showing results of prediction using actual observation results.

FIG. 11 is a graph showing a result of prediction using actual observation results.

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Kenichi Yoshihara             2-9-26 Yamane, Zushi City, Kanagawa Prefecture F-term (reference) 5B056 BB01 BB02 BB62 HH00

Claims (5)

[Claims] 1. A short-term prediction system for predicting fluctuations of a prediction period based on observation results of a plurality of past observation periods for a phenomenon that changes continuously for a short period of time. the observation period _{(-mt 0 <t i ≦ 0} :
Let S be the observation result of a set of observation result Q of the observation result Q of the time period corresponding to the prediction period and the prediction period of t ₀ )
Observations of S) for each unit time t ₀ of the matrix X (t
_i )) storage means for storing the row vector X for each S of the stored observation results
The average of the corresponding unit time of _α (t _i ) To find the means of calculating and the relationship of the difference between the obtained average value and each observation result, And a simultaneous equation for correlating W _{i, j} with respect to the time t _k of the prediction period Solve, respective coefficients _{A 0,} A 1, · · ·, means for determining the _{A m,} using a solution of the simultaneous equations, the value at time _{t k} of the prediction period Y _{(t k)} Equation 4 The In the short-term prediction system, the obtained Y (t _k ) is output as a prediction value in the prediction period. 2. The short-term prediction system according to claim 1, wherein the observation result is at least one state of meteorology, sea state, and geography observed at a plurality of observation points where observation is regularly performed. A short-term prediction system characterized by being 3. A recording medium that stores a program that enables a computer system to implement the short-term prediction system according to claim 1. 4. A program capable of causing a computer system to implement the short-term prediction system according to claim 1 or 2. 5. A short-term prediction method for predicting fluctuations in a prediction period based on observation results of a plurality of past observation periods for a phenomenon that changes continuously for a short period of time. the observation period _{(-mt 0 <t i ≦ 0} :
Let S be the observation result of a set of observation result Q of the observation result Q of the time period corresponding to the prediction period and the prediction period of t ₀ )
Observations of S) for each unit time t ₀ of the matrix X (t
_i ), the row vector X for each S of the stored observation results
The corresponding average of _α (t _i ) per unit time To calculate the relationship between the calculated mean value and the difference between each observation result and [Equation 6] And the simultaneous equations that correlate W _{i, j} for the time t _k of the prediction period Solve, respective coefficients _{A 0,} A 1, · · ·, determining a _{A m,} using a solution of the simultaneous equations, Equation 8] The value Y _{(t k)} at time _{t k} of the prediction period And a step of obtaining the obtained Y (t _k ) as a prediction value in a prediction period.