JP2010020442A - Method and device for creating prediction scenario of future demand regarding transaction object - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method of creating a plurality of prediction scenarios to revise the future prediction results only from the past prediction results by a current prediction model and the past records regardless of the prediction model and to analyze the future prediction results considering uncertainty. <P>SOLUTION: M-fold time series of errors containing N elements are calculated by subtracting the time series of recorded values from the time series of predicted values for each variation period using the time series data which are the predicted values of a past predetermined period of the periodically varying future demand and the time series data which are the recorded values of a past predetermined period over M times of past variation period. The variance of the data containing M elements obtained by taking out only the predetermined elements of the time series of errors and the covariance of M data consisting only of the i-th element of the time series of errors and M data consisting only of the predetermined elements are determined. Using a variance/covariance matrix containing N×N elements created from the above results, multi-variant joint probability distribution and sets of random numbers are obtained, and many prediction scenarios are created from the sets of the random numbers. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a method for creating a forecast scenario for future demand related to a transaction object and a creation apparatus thereof.

  Predicting the future demand related to the transaction object, that is, the future demand amount of electric power, the inventory amount of goods, etc. is very important in preparing the power generation plan and the sales plan. In particular, in the case of electricity demand, it is necessary to predict the maximum annual demand in order to formulate a capital investment plan, and to predict the future demand in the next day in order to create a generator operation plan. is there. Then, the future demand is predicted using multiple regression analysis or the like based on the past actual demand forecast results or the predicted temperature value.

  In addition, there are many fields that need to be predicted in the future, such as future demand and market prices for tangible and intangible products, market prices for financial products such as stocks, interest rates, exchange rates, and other economic indicators. For these predictions, prediction models using mathematical models have been developed and put into practical use in the respective fields.

  The purpose of making these predictions is to use the prediction results directly for some actions (such as buying or selling the product), but to perform some analysis based on the prediction results and use the results. There are also many. For example, if it is future demand, create a power plant operation plan based on time-series data, which is the change in future demand for the day, calculate the amount of fuel required, and calculate income from power supply. Can be.

  These amounts depend on the time-series data path that changes every moment, and are calculated as, for example, an integral value of a function of the future demand amount at each time. For this purpose, it is not sufficient to predict the maximum demand amount of the day alone, and it is necessary to predict all future demand amounts at each time (for example, 24 points per hour).

  On the other hand, the prediction value usually has an error, and it is useful to know the prediction error in advance. For this reason, various methods for evaluating the prediction error have been devised. For example, for future demand, there are a case where the error of the predicted value of the maximum demand on a certain day is obtained, and a case where the error of each time of the day is obtained by 24 points. In the latter case, the error at each time is usually related to each other, and it is not sufficient to obtain the error at each time independently and arrange 24 pieces, and it is necessary to evaluate 24 errors together. Therefore, the latter is much more difficult.

  As technologies in such a field, there are those shown in Patent Document 1 and Patent Document 2. First, in Patent Document 1, future demand for a predetermined period in the future is predicted based on weather information, but in order to improve the accuracy of prediction, the predicted value is corrected every moment using the latest weather data. Further, in order to evaluate a prediction error of a future demand prediction value at a specific time, a standard deviation between a demand prediction error and a weather prediction error is used.

  Also in Patent Document 2, future total power demand is predicted from temperature and humidity data and past future demand data. This case is a prediction of the total power demand for one day, but is characterized in that the error is corrected using a technique such as a neural network. That is, when the error has a specific tendency, such as when the temperature fluctuates greatly, it is found by a method such as a neural network and the error is corrected.

  As described above, the conventional technology usually aims to reduce the error by various devices. However, when considering the error in the predicted value for a predetermined period, a simple sum or average of the error in the predicted value at each time does not mean the overall error.

  The error may have a certain tendency, and in such a case, the error has a series correlation. In the case where the error is large or there is a series correlation (a certain tendency) in the error, the conventional technique is that the prediction model is usually corrected.

  However, it is impossible to create a completely ideal prediction model, and it is generally very difficult to improve the prediction accuracy in fields that truly require prediction, such as social phenomena and stochastic phenomena. There is a problem. In these fields, the underlying factors change with time, and it is often necessary to frequently change the prediction model accordingly.

  On the other hand, it is also important to verify the accuracy of the prediction model by comparing the predicted result with the actual result. In this case, there is a problem that if the prediction model is frequently corrected, it is difficult to verify the validity of the prediction result.

  FIG. 13 is a conceptual diagram showing a conventional prediction method. In order to predict future values from information obtained so far, a large number of models such as regression models and neural networks are used.

  In this case, a past actual value may be used, but usually a past predicted value is not used. In addition, error bars, standard deviations, and the like are used to evaluate future predicted value errors.

  FIG. 14 shows the concept of conventional prediction results and error evaluation. In FIG. 14, a prediction error bar is drawn at each time. As an example of the error bar, the standard deviation of error may be used. However, the above-described scenario cannot be taken into account in such a method that evaluates the prediction error at each time.

  As mentioned above, it is difficult to create a highly accurate prediction model with the prior art, and even if a highly accurate prediction model can be created, the prediction model becomes inapplicable at different times. There's a problem.

  Further, since the standard deviation is used for the error evaluation, the numerical error can be evaluated, but it is difficult to generate time series data in consideration of the error.

  Therefore, the idea that not only the accuracy of the prediction model but also the means of the prediction model is not a problem has appeared. Rather, it is based on the premise that there is an error in the prediction, and the future phenomenon is analyzed using a plurality of prediction results.

  In addition, in the case of prediction of time series data as described above, it is necessary to perform prediction as a series of time series instead of prediction at each time, and the error is not an error at each time, but a plurality of possible times. It is necessary to perform prediction and error evaluation as a collection (ensemble) of series data.

  An example of a prediction method that uses a plurality of prediction results is the concept of ensemble prediction. For example, the prediction result obtained by using the average value of a plurality of prediction results is approximated. Patents related to such a technology include, for example, Patent Document 3.

According to Patent Document 3, in the weather simulation necessary for weather prediction, even if the initial value is changed slightly, the result is greatly different and prediction becomes difficult. For this reason, the accuracy of prediction is improved by averaging a plurality of simulation results having different initial values.
JP 2007-47996 A JP 5-18995 A Japanese Patent No. 3904420

  Here, each scenario in Patent Document 3 compares a low-pressure movement vector with a plurality of predicted movement vectors and selects a prediction result, and predicts and evaluates errors as an aggregate of a plurality of time-series data. It is carried out.

  However, it is solved as an initial value problem of a specific physical equation, and the prediction model is only specified.

  The present invention has been made in consideration of the above-mentioned points. The present invention is not concerned with the prediction model, the future prediction result is corrected based on only the past prediction result and the past actual value based on the current prediction model, and is uncertain. It is an object of the present invention to provide a method and apparatus for creating a plurality of prediction scenarios capable of analyzing future prediction results in consideration of characteristics.

In order to achieve the above object, in the present invention,
In a method for creating a scenario for forecasting future demand for a transaction object based on actual values in a predetermined period in the past,
Time-series data including N elements that are predicted values of a past predetermined period of time-varying future demand, and time-series data including N elements that are actual values of the past predetermined period , Store over the past M fluctuation cycles,
Subtracting the time series of the predicted values and the time series of the actual values for each of the M fluctuation periods, and calculating M time series of errors including the N elements,
Calculating a variance of data including M elements obtained by extracting only the k-th element of the error time series;
The covariance between M data consisting only of the i-th element of the error time series and M data consisting only of the j-th element is all (i, j) where i ≠ j. Calculate over a set of
Create a variance-covariance matrix containing N × N elements from the results obtained by calculation,
Calculate a multivariate joint probability distribution using the variance-covariance matrix;
Generating a set of random numbers including N elements K times based on the multivariate joint probability distribution;
A forecast scenario creation method for future demand for a transaction object, characterized in that K forecast scenarios are created from the set of random numbers,
And a device for creating a scenario for forecasting future demand for a transaction object based on actual values for a predetermined period in the past,
Time-series data including N elements that are predicted values of a past predetermined period of time-varying future demand, and time-series data including N elements that are actual values of the past predetermined period Means for storing over the past M fluctuation cycles;
Means for subtracting the time series of the actual values from the time series of the predicted values and calculating M time series of errors including the N elements for each of the M fluctuation periods;
Means for calculating a variance of data including M elements obtained by extracting only the k-th element of the error time series;
The covariance of M data consisting only of the i-th element and M data consisting only of the j-th element in the time series of the errors is all (i, j) where i ≠ j. Means for calculating over a set of
Create a variance-covariance matrix containing N × N elements from the results obtained by calculation,
Means for calculating a multivariate joint probability distribution using the variance-covariance matrix;
Means for generating a set of random numbers including N elements K times based on the multivariate joint probability distribution;
And a means for creating K prediction scenarios from the set of random numbers.
However, N, M, i, j, and k are all positive integers.
Is to provide.

  As described above, the present invention obtains a plurality of time-series data obtained by adding a time series of errors to each of a plurality of time-series data of past predicted values and actual values related to future demand, and the distribution and sharing of the time-series data. Since a set of multivariate joint probability distributions and random numbers is obtained by creating a matrix by calculating variance, and a number of forecast scenarios according to the random number pairs are created, the demand is predicted using the scenario density. An easy prediction scenario can be provided.

Basic concept of invention

  The present invention adopts the concept of analyzing future phenomena using a plurality of prediction results on the assumption that there is an error in prediction. The plurality of prediction results are referred to as “prediction scenario”.

  The present invention predicts a series of time series of data such as future demand in a predetermined period in the future as a plurality of scenarios, and sets the operation patterns of the generator and related devices in that period to a plurality of scenarios. It is related to a system for predicting the total demand and sales for that period, and the operation plan of the generator and related devices, including uncertainties, and has the following characteristics.

  Next, the outline of the present invention will be described with reference to FIGS.

Summary of the present invention

  FIG. 1 is a conceptual diagram showing a prediction method of the present invention. In this case, typically only the past predicted value and the past actual value are used. In order to obtain past prediction values, some kind of prediction model A has been used in the past, but the method of this prediction model A may be anything in the present invention.

  Past error data is obtained from the difference between the past actual value and the past predicted value. This is calculated every predetermined period such as the past day, week or month, and sets of error data are obtained for the number of days, weeks or months for which data exists. It is assumed that there are K error data sets (K ≧ 2).

  On the other hand, the future predicted value is predicted by, for example, the prediction model A. The present invention is characterized in that the error of the future prediction result is evaluated using the above-described error data set.

  That is, the error at each time is calculated from a set of K error data, and the correlation between the errors at each time is calculated, not only the variance of the error at each time (square of the standard deviation) but also each The error covariance between times is calculated to obtain the error relationship between each time.

  In the following, the case where the prediction cycle is one day, the time interval for prediction is one hour, and each error data has 24 errors will be specifically described as an example. From the set of K error data, K errors are collected at each time, and a variance vi of K error values at the i-th time is calculated.

  When this is performed at each time, 24 pieces of distributed data (i = 1 to 24) are obtained. In order to calculate the variance, K needs to be at least 2 or more. Next, the covariance between each time is calculated. This is the covariance v1,2 of the K error values at time 1 and the K error values at time 2, and the covariance v1,2 of the K error values at time 1 and the K error values at time 3. 3, ... Covariance v23,24 of K error values at time 23 and K error values at time 24 is calculated.

  In this case, 276 covariances, which are the number of combinations for selecting two sets without allowing duplication from 24 times, are obtained. The characteristics of error data are determined by these variances and covariances.

  Next, a method for creating a prediction scenario of the present invention will be described. The error at each time is considered to follow a normal distribution with an average of 0 and a variance of vi if the prediction model is ideal. Therefore, in the conventional method, the error is evaluated with the standard deviation σi = (vi) 0.5, for example.

  On the other hand, in the present invention, in order to evaluate an error as a scenario, not only the variance (or standard deviation) of each time but also the covariance between each time is used. In order to create an error scenario from this covariance data, for example, an N-dimensional normal distribution is used. The N-dimensional normal distribution is a distribution according to a probability density function as shown in Equation 1 below.

ここで、Tは転置べクトルを示し、Σ^-1はΣの逆行列を示す。
と表される。また、Σは、下記式２すなわち、

ここで、var(x_i)はx_iの分散を、cov(x_i,x_j)はx_iとx_jの共分散を表す。すなわち、前述の例では、var(x_i)は時刻iの誤差Ｋ個の分散を、cov(x_i,x_j)は時刻iの誤差Ｋ個と時刻jの誤差Ｋ個の共分散を表す。
で計算される分散共分散行列である。 When an N-dimensional vector x = (x1, x2, x3, ..., xN) follows a multivariate normal distribution, the probability density function that x satisfies is the expected value vector of x μ = (μ ₁ , μ ₂ , μ ₃ , ..., μ _N ) Using the following formula 1, that is,

Here, T represents a transposed vector, and Σ ⁻¹ represents an inverse matrix of Σ.
It is expressed. In addition, Σ is the following formula 2, that is,

Here, the variance of var (x _i) is x _i, cov (x _i, x _j) represents the covariance of x _i and x _j. That is, in the above example, var (x _i ) represents the variance of error K at time i, and cov (x _i , x _j ) represents the covariance of error K at time i and error K at time j. .
Is the variance-covariance matrix calculated by

  In the present invention, for example, a large number of random numbers (N-dimensional normal random numbers) according to an N-dimensional normal distribution are generated, and this set is regarded as time series data to constitute an error scenario. In the above example, N = 24. In the N-dimensional normal random number, each random number is related to each other, and this relationship is expressed by the variance-covariance matrix of the above equation 2.

  By generating a set of M random numbers by the above method, M error scenarios are obtained. The value of M is completely arbitrary and does not depend on the value of K or the value of N. For example, one million scenarios can be generated even if there are only two days of past performance data.

  By adding M error scenarios and future prediction values predicted by the original prediction model, M prediction scenarios are obtained. Eventually, M + 1 prediction scenarios are obtained together with the prediction values.

  Note that the probability density distribution for generating random numbers need not be limited to a normal distribution. For example, an exponential distribution, a Poisson distribution, a log normal distribution, a Weibull distribution, or the like can be used. Normally, a normal distribution is used. However, when past error data is examined, when it is determined that there is a clear deviation from the normal distribution, it is desirable to use a probability distribution that can approximate the distribution.

  FIG. 2 is an image diagram of the scenario obtained as described above. A bold line 1 in the figure is a prediction result obtained by some prediction method, and corresponds to a prediction value by a conventional method. The prediction method is preferably the same as the prediction method used in the past, but is not necessarily the same. A large number of thin lines 3 drawn in FIG. 2 are prediction scenarios.

  Although only five thin lines are drawn in FIG. 2, in practice, a very large number of scenarios can be easily created. Moreover, the number density in the vicinity of the line of this prediction scenario is proportional to the probability of occurrence of the line.

  Therefore, for example, if this is a forecast scenario of future demand (sales), the expected value of sales can be calculated by multiplying the power price at each time and summing up and dividing by the number of scenarios for all forecast scenarios. .

  When the power price is constant, if the average error is 0, the sales according to the conventional method coincide with the sales according to the present invention, but the result differs when the price is a function of demand.

Also, even if the price is constant, if the prediction model is not accurate (when the error average is not 0), the value calculated by the conventional method is different from the average value calculated by the above method, By using the present invention, a more accurate average value can be calculated. If the prediction model is ideal, the error between a number of predicted demand patterns and the corresponding actual value will be zero at each time.

  On the other hand, the standard deviation of the error at each time is not always constant. For example, it is considered that the standard deviation in the daytime is large and takes a small value at night. Furthermore, the error at each time is related to the error between the previous and subsequent times.

  For example, if the prediction at 1:00 pm shifts to the plus side, the prediction at 2:00 pm is also considered to shift to the plus side. This is because there is a high possibility that the event that caused the prediction at 1 pm to deviate (eg, the temperature was higher than the prediction) continued at 2 pm, rather than a defect in the prediction model.

  Each scenario depicted by one thin line in FIG. 2 has a meaning, and is different from a simple arrangement of errors. Therefore, most prediction scenarios may concentrate on the upper side of the predicted value 1 or on the lower side. In some cases, as shown by a dotted line 4 in the figure, a scenario may be considered in which the lower side of the predicted value is initially passed and the upper side of the predicted value is passed in the latter half.

  In the present invention, from the relationship between past performance data and predicted values, if such a possibility actually exists, by creating a very large number of scenarios, such a scenario can be reproduced with the correct probability. Is done. That is, even if a very ideal prediction model is used, a series correlation occurs in the error, and it is not necessary to correct the prediction model just because a series correlation exists.

  FIG. 3 is an example of calculating an error scenario according to the present invention. Fig. 3-a) shows the prediction results and actual values for the past five days. Fig. 3-b) shows the error data for 5 days calculated from the prediction result and the actual value. Fig. 3-c) shows the results for five days superimposed. Fig. 3-d) shows the predicted values for 5 days superimposed. Fig. 3-e) shows the error data for 5 days superimposed. FIG. 3F shows an example in which a variance-covariance matrix is calculated from the error data shown in FIG. 3-E) and 30 sets of random numbers are generated to create 30 scenarios.

によって計算される量であり、−1.0〜１.0の値を取る。例えば、時刻１の誤差が常に正であるとき、時刻２の誤差も常に正であれば、ρ（x1,x2）は１に近い数値となる。時刻１の誤差と時刻２の誤差とが独立であれば、ρ（x1,x2）は０になる。i=jの場合には、自分自身の相関係数であるから常に１となる。 In order to see the relationship between error data, the correlation coefficient is easier to understand than the variance or covariance. Since each term of the variance-covariance matrix is a value of variance and covariance, it is an amount corresponding to the degree of dispersion of raw data. On the other hand, the correlation coefficient ρ (xi, xj) is expressed by the following Equation 3,

Which is a quantity calculated from -1.0 to 1.0. For example, when the error at time 1 is always positive and the error at time 2 is always positive, ρ (x1, x2) is a value close to 1. If the error at time 1 and the error at time 2 are independent, ρ (x1, x2) is zero. In the case of i = j, it is always 1 because it is the correlation coefficient of its own.

  FIG. 4 shows an example of the prediction scenario finally obtained. In this case, 30 scenarios are created, but the number of scenarios is not limited. In this case, the scenarios are concentrated in the vicinity of the original predicted value, and the original predicted value is the most probable scenario. The density of the number of scenarios near a scenario is proportional to the probability that the scenario will occur. By increasing the number of scenarios, more accurate prediction can be performed.

  FIG. 5 is a correlation coefficient matrix in the case of FIG. Naturally, the diagonal terms of the matrix are all 1, but the correlation coefficients of adjacent terms such as the (1,1) term and the (1,2) term are relatively high, (1 , 24), the correlation coefficient of the data at locations separated in time is relatively small.

  Further, in the case of future demand, the correlation between daytime data errors as in (12, 13) is compared with late-night data as in (1, 2) and (23, 24). Is higher. In FIG. 5, these are all “1.00” because the number of digits is small. Strictly speaking, they are very close to 1 but not 1.

  As a result of predicting the time-series data of the future demand of the day, the maximum value does not change so much, but there may be an error in the prediction because the rise time of the demand is different.

  FIG. 6 shows an example of such a case. FIG. 7a) shows the predicted values and actual values for the past three days. In this case, it can be seen that the rise time of demand is slightly changed. FIG. 7b) is a demand scenario created by the method of the present invention. It can be seen that a scenario has been created that captures the characteristics of data that errors are likely to occur at the rise time.

  Hereinafter, an embodiment of the present invention will be described with reference to FIGS.

  FIG. 8 shows a first embodiment of the present invention. In the first embodiment, first, K sets of error data 104 are created by the error calculation function 103 from N pieces of actual data 101 and N pieces of prediction data 102 for the past K periods, and the variance covariance calculation function 105 is used to create an N × N variance-covariance matrix 106 from K pieces of data at each time of 1 to N and give it to the random number generation function 111.

  The random number generation function 111 is provided with M number of scenarios 109 and a probability density function 110, and the generated random number is supplied to the prediction scenario creation function 112. The prediction scenario creation function 113 is provided with prediction data 114 based on the prediction method 115 and creates (M + 1) sets of prediction scenarios 116. In this case, the prediction value created by the original prediction model is added to create M + 1 scenarios, but it is not always necessary to add the original prediction value.

  FIG. 9 shows a second embodiment of the present invention. The second embodiment is an example in which an error model is created as preprocessing in addition to the first embodiment shown in FIG.

  For this purpose, the data selection function 124 selects at least one of the weather data 121, the event data 122, and the day type data 123, and adds the past error data 103 A obtained by the error calculation function 103 to the error data 104. To give. Others are the same as the first embodiment.

  FIG. 10 shows an example in which both demand and price prediction scenarios are created simultaneously. Here, demand and price are described as an example of a plurality of prediction targets. For example, in the electricity market, both demand and price volumes may fluctuate and both may need to be predicted simultaneously.

Even in such a case, the present invention can be applied as it is.

  That is, as shown in FIG. 10, demand error data and price error data are created from past actual values and predicted values of demand and price. At this time, it is assumed that error data is obtained at times 1 to N for demand and times 1 to N ′ for prices.

  Next, the N + N ′ pieces of data are regarded as a continuous time series, and similarly to the case of creating a demand-only forecast scenario, a (N + N ′) × (N + N ′) variance-covariance matrix is created, M (N + N ′) random number sets can be generated. In this way, a demand scenario and a price scenario can be created simultaneously.

The above prediction scenario creation method is not limited to a two-dimensional function. For example, Equation 4 below:
[Formula 4] Demand = X (time)
Price = Y (time)
Temperature = Z (time)
It is assumed that these prediction models are created. In this case, a future prediction scenario is expressed as a trajectory in a three-dimensional space.

  FIG. 11 shows a third embodiment of the present invention, which is an example of displaying a scenario to be predicted in a three-dimensional space.

  The number of variables in the parentheses on the right side of Equation 4 need not be one. For example, assuming that the demand is a function having time and day of the week as variables, a scenario in a three-dimensional space can be considered similarly.

  Furthermore, these data need not be continuous in time. The price of many products may be arranged in a line, or the highest temperature of the prefectural offices nationwide. As long as a certain prediction model and past actual value and prediction value data exist, it is applicable to error estimation and scenario creation of any prediction model.

  In general, when a prediction target is represented by a set of P data, from these prediction models and the past actual values and prediction results, by considering N as P in the above-described method, Many sets of future predicted values of P data can be created.

  After the scenario is created, an evaluation value can be obtained by performing some calculation for each scenario. The evaluation value may be anything such as the amount of sales for each scenario, the cost, the amount of necessary materials, the probability of an accident, and the like.

  FIG. 12 shows an example in which the prediction result can be easily understood visually. That is, an important scenario can be discriminated by arranging and displaying the evaluation values in descending order or decreasing order. In this case, the prediction result can be easily understood visually by displaying not only a table but also a bar graph or the like.

  It is also possible to evaluate the probability of occurrence of several scenarios and arrange them in the order of occurrence probability. Furthermore, according to the present invention, cumulative probabilities and Pareto charts in which these scenarios occur can be created and provided for the scenario creator to analyze the prediction results.

The conceptual diagram which shows the prediction method of this invention. The conceptual diagram of the prediction scenario by this invention. The characteristic view which shows the example of the calculation result of an error scenario. The characteristic view which shows the example of the prediction scenario finally obtained. Explanatory drawing which shows the correlation coefficient matrix in the case of FIG. The characteristic view which shows the prediction error of the maximum value, and the prediction error of rise time. FIG. 7A is a characteristic diagram showing an example of creating a scenario when there is an error in the rise time, FIG. 7A shows a past actual value and a predicted value, and FIG. 7B shows a predicted scenario. The block diagram which shows the 1st Example of this invention. The block diagram which shows the 2nd Example of this invention. Explanatory drawing which shows the example which created the forecast scenario of both the demand and the price simultaneously. Explanatory drawing which shows the example which created the scenario of the three-dimensional space in the 3rd Example of this invention. It is explanatory drawing which shows the example which calculated the evaluation value of each scenario, and displayed on the graph and the table | surface, FIG.12 (a) shows the example displayed in order of the scenario number, FIG.12 (b) displays in order with large evaluation value An example is shown. Explanatory drawing which shows the example of the conventional prediction method. The conceptual diagram of the prediction result and error evaluation by a conventional method.

Explanation of symbols

1: Prediction value 2: Error bar 3: Prediction scenario 4: Other examples of prediction scenario 101: Past performance data 102: Past prediction data 103: Error calculation function 104: Error data 105: Variance covariance calculation function 106: Covariance matrix 111: Random number generation function 112: Error scenario creation function 113: Prediction scenario creation function 114: Prediction data 115: Prediction method 116: Prediction scenario 121: Weather data 122: Event data 123: Day type data 124: Data selection function

Claims (9)

In a method for creating a scenario for forecasting future demand related to a transaction object based on actual values for a predetermined period in the past,
Time-series data including N elements that are predicted values of a past predetermined period of time-varying future demand, and time-series data including N elements that are actual values of the past predetermined period , Store over the past M fluctuation cycles,
For each of the M fluctuation cycles, subtract the time series of the actual values from the time series of the predicted values to calculate M time series of errors including the N elements,
Calculating a variance of data including M elements obtained by extracting only the k-th element of the error time series;
Covariances of M data consisting only of the i-th element and M data consisting only of the j-th element in the time series of errors are all (i, j) where i ≠ j. Calculate over a set of
Create a variance-covariance matrix containing N × N elements from the results obtained by calculation,
Calculate a multivariate joint probability distribution using the variance-covariance matrix;
Generating a set of random numbers including N elements K times based on the multivariate joint probability distribution;
A forecast scenario creation method for future demand for a transaction object, characterized in that K forecast scenarios are created from the set of random numbers.
However, N, M, i, j, and k are all positive integers. In the method for creating a scenario for forecasting future demand for a transaction object according to claim 1,
For each forecast scenario created, calculate the amount depending on each scenario based on the given formula, compare the above amount for each scenario,
A method for creating a scenario for forecasting future demand related to a transaction object, wherein the scenario is displayed in a table, graph, histogram or the like in order of scenario number, ascending order or descending order. In the method for creating a scenario for forecasting future demand for a transaction object according to claim 1,
A method for creating a scenario for forecasting future demand for a trading object, wherein a multivariate normal distribution is used as the multivariate simultaneous probability distribution. In the method for creating a scenario for forecasting future demand for a transaction object according to claim 1,
When creating error data, a transaction characterized by selecting common date data based on the forecast period and the same day of the week, the same month or the same year, or a combination of these, and creating error data How to create forecast scenarios for future demand for objects. In the method for creating a scenario for forecasting future demand for a transaction object according to claim 1,
When creating error data, select the data for which the difference between the forecast period and one or more of the maximum temperature, average temperature, or minimum temperature value within the period is less than the specified value. A method for creating a scenario for forecasting future demand for a transaction object, characterized by creating data. In the method for creating a scenario for forecasting future demand for a transaction object according to claim 1,
Create forecast scenarios for each of the two or more forecast targets from two or more forecast targets, two or more forecast models corresponding to each, and past actual values and forecast results. A method for creating a scenario for forecasting future demand for a characteristic transaction object. In the method for creating a scenario for forecasting future demand for a transaction object according to claim 1,
When the prediction target is composed of a set of L numerical values and each prediction target is defined by some variables, a prediction scenario of each transaction target for each prediction target is created as a trajectory in the L-dimensional space. A method for creating a scenario for forecasting future demand for a transaction object. In the method for creating a scenario for forecasting future demand for a transaction object according to claim 1,
When a prediction target is expressed by a set of P data, a time series including N elements is regarded as a set of P data from these prediction models, past actual values, and prediction results. A method for creating a predicted value of a future demand for a transaction object, wherein one or more pairs of prediction values of the transaction object of P pieces of data are created. In a device that creates a scenario for forecasting future demand related to a transaction object based on actual values for a predetermined period in the past,
Time-series data including N elements that are predicted values of a past predetermined period of time-varying future demand, and time-series data including N elements that are actual values of the past predetermined period Means for storing over the past M fluctuation cycles;
Means for subtracting the time series of the predicted values and the time series of the actual values and calculating M time series of errors including the N elements for each of the M fluctuation periods;
Means for calculating a variance of data including M elements obtained by extracting only the kth element of the time series of the error;
The covariance of M data consisting only of the i-th element and M data consisting only of the j-th element in the time series of the error is represented by all (i, j) where i ≠ j. Means to calculate across the set;
Create a variance-covariance matrix containing N × N elements from the results obtained by calculation,
Means for calculating a multivariate joint probability distribution using the variance-covariance matrix;
Means for generating a set of random numbers including N elements K times based on the multivariate joint probability distribution;
Means for creating K prediction scenarios from the set of random numbers; and a forecast scenario creating device for future demand related to a transaction object.
However, N, M, i, j, and k are all positive integers.

Images