JP2004086896A - Method and system for constructing adaptive prediction model - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method or a system for constructing an adaptive prediction model which makes efficient construction of a model with high accuracy possible by using time-series data. <P>SOLUTION: When the error of a predicted value outputted from a prediction model becomes large, that is, the characteristics of time-series data which the prediction model learns become different from time-series data at the point in time of prediction, the prediction model is learned again and the model is updated. When the error becomes still larger, that is, overall review of the prediction model becomes inevitable, the construction of the prediction model itself is changed to revise the entire prediction model. <P>COPYRIGHT: (C)2004,JPO

Description

The present invention, the power demand, water demand, heat demand and other demand, various sales, or time-series data such as various economic indicators, or, using other relevant information, the predicted value that is a future value The present invention relates to an adaptive prediction model construction method and an adaptive prediction model construction system for constructing a prediction model to be calculated.

Generally, prediction of time-series data is performed using the periodicity inherent in the time-series data, correlation with related information, and the like. For example, in the case of time-series data on a daily basis, the correlation often differs depending on the difference between weekdays and holidays, the season, and the like.

Therefore, a method is adopted in which a plurality of prediction models are constructed in advance by dividing the period for each feature or each section, and these prediction models are switched and applied at the time of prediction. For example, a prediction system described in Non-Patent Document 1 is cited. In this document, six types of prediction models for spring, spring / summer, summer, summer / fall, autumn, and winter are constructed, and a prediction method / prediction is used in which these prediction models are switched and used. A system is disclosed.

In addition, a method of performing prediction while updating the prediction model periodically (daily, weekly, monthly, etc.) has been performed.
In any of the above methods, a period during which the prediction model is applied and an update timing of the prediction model are determined before application.
In the prediction model of Non-Patent Document 1 cited earlier, the application period is set to be from April 1 to May 31 in spring, from June 1 to July 31 in spring and summer, and from August 1 to summer in summer. August 31, summer and autumn are defined in advance from September 1 to September 30, autumn is defined as October 1 to October 31, and winter is defined as November 1 to March 31. .

In addition, various models are used as a prediction model for performing such a prediction, and a multiple regression equation or a neural network is generally used as a prediction model using a correlation.
The multiple regression equation is used when it is necessary to explain an output reason to an operator or an operator.
Further, the neural network can construct a nonlinear prediction model, and is used in a case where the prediction accuracy is important.

Here, prediction modeling of a neural network, which is a specific example of modeling, will be described with reference to the drawings. FIG. 9 is a conceptual diagram illustrating a multilayer neural network. In general, a neural network has a multilayer neural network structure including an input layer, an intermediate layer, and an output layer, as shown in FIG. 9, and further, elements are provided in the input layer, the intermediate layer, and the output layer. It has coupling between the element between the layer and the intermediate layer and between the element between the intermediate layer and the output layer. In this neural network, the elements in the input layer correspond to the input factors, and the elements in the output layer correspond to the output factors.

The coupling coefficient is a coefficient for representing the weight of coupling between elements of the neural network. If the coupling coefficient is large, the coupling is weighted, that is, it is a required coupling, and if the coupling coefficient is small, the coupling is not weighted, that is, it is an unnecessary coupling. .
Such prediction modeling of a neural network means that a desired output value is obtained from an output layer element (output factor) for an input value (time series data) input to a plurality of input layer elements (input factors). This means changing the coupling coefficient between the input layer and the intermediate layer and between the intermediate layer and the output layer.

In such a prediction model, a change in the coupling coefficient and a change in the number of elements in the input layer / intermediate layer correspond to the update of the prediction model itself.
Further, a change in the number of neural networks used corresponds to a change in the prediction model configuration. This is, for example, a change in whether to use four models for spring, summer, autumn, and winter, or to use three models for summer, winter, and middle season.

IEEJ Transactions on Electronics B, Power and Energy Division Magazine, Vol. 120-B, no. 12, pp 1550-1556, (2000), "Development of NN applied power demand forecasting system"

In the above-described conventional technology, the update timing and the application period of the prediction model are determined in advance.
(1) A method of constructing a prediction model for each period or section having different characteristics,
(2) There is a method of predicting while periodically updating the prediction model. However, these methods have problems.

(1) In the method in which a prediction model is constructed for each period or section having different characteristics, accurate prediction is impossible if the time-series data to be predicted shows an unexpected tendency. In addition, when the characteristics of a certain period or section change, it is difficult to perform sufficiently accurate modeling in a fixed period or section, and the prediction error may increase.

(2) In the method of performing prediction while updating the prediction model periodically, even when the prediction model is not required to be updated, the prediction model is updated and an unnecessary calculation load is imposed on the system. .

The present invention has been made in order to solve the above problems, and a first object of the present invention is to provide a system in which time series data to be predicted shows an unexpected tendency without imposing an unnecessary load on the system. Another object of the present invention is to provide an adaptive prediction model construction method and an adaptive prediction model construction system that enable highly accurate prediction.

Further, the second object of the present invention is to provide a method for dividing time-series data, or a method for constructing a prediction model without prior knowledge. It is an object of the present invention to provide an adaptive prediction model construction method / adaptive prediction model construction system which automatically determines a suitable prediction model configuration and enables highly accurate prediction.

Generally, there is a case where the user does not know how to construct a prediction model but has knowledge about the characteristics of time-series data. For example, in the case of power demand data, the demand amounts are significantly different on weekdays, Saturdays, and holidays. When the temperature is high and when the temperature is low, there is a demand for cooling and heating, so the demand increases. At medium temperatures, demand is small. In terms of the seasonal division of spring, summer, autumn and winter rather than temperature, demand is large in summer and winter, but demand is small in spring and autumn. If the prediction model is configured based on such empirical knowledge of the characteristics of the time-series data, the prediction model can be easily understood by the user.

A third object of the present invention is to make it possible to perform high-precision prediction while automatically adjusting a prediction model configuration by using the knowledge as an initial state when there is prior knowledge about the target time-series data. To provide an adaptive prediction model construction method and an adaptive prediction model construction system.

Generally, it is an object of the present invention to provide an adaptive prediction model construction method or an adaptive prediction model construction system that enables efficient construction of an accurate model using time-series data.

An adaptive prediction model building method according to the invention described in claim 1 has been made to solve the above problem.
In an adaptive prediction model construction method for adaptively constructing a prediction model for predicting future time series data based on past time series data,
A prediction model construction step of constructing a prediction model using past time series data,
Performing a prediction using the prediction model, a prediction execution step of obtaining a predicted value,
A prediction value correction step of calculating a correction coefficient or a correction amount for correcting the prediction value based on a past prediction error or a model error,
Forecast error, model error, based on the correction coefficient or the correction amount, whether the model needs to be updated, or whether the model configuration needs to be updated, a prediction model update determination step for determining
When it is determined that model update and model configuration update are unnecessary, a predicted value correction execution step of correcting the predicted value,
A prediction model update step of updating the prediction model using the newly accumulated time-series data when it is determined that the model update is necessary;
A prediction model configuration update step of updating the prediction model configuration using the newly accumulated time-series data when it is determined that the model configuration needs to be updated;
It is characterized by having.

Further, an adaptive prediction model building method according to the second aspect of the present invention includes:
The adaptive prediction model building method according to claim 1,
When the prediction error, the model error, the correction coefficient or the correction amount exceeds the allowable value for model update determination, it is determined that the prediction model needs to be updated, and the tendency / characteristics of the time series data to be predicted change. Alarm output step to output an alarm to notify that
It is characterized by having.

Further, an adaptive prediction model building method according to the third aspect of the present invention,
The adaptive prediction model building method according to claim 1,
When the prediction error, the model error, the correction coefficient, or the correction amount exceeds the allowable value for model configuration update determination, it is determined that the prediction model configuration needs to be updated, and the trend / characteristics of the time series data to be predicted changes. Alarm output step to output an alarm to notify that
It is characterized by having.

Further, the adaptive prediction model building method according to the invention described in claim 4 is characterized in that:
In the adaptive prediction model building method according to any one of claims 1 to 3,
The prediction model construction step and the prediction model configuration update step,
Analyzing time-series data of multiple factors, finding an input factor having an input-output relationship with a certain output factor from among the multiple factors, and modeling the input-output relationship,
A preliminary clustering step of performing clustering using all of the time series data to be used and dividing the cluster into preliminary clusters;
A clustering factor selection step of calculating the importance of the clustering factor for each preliminary cluster and selecting a clustering factor;
A clustering step of performing clustering using the time-series data related to the clustering factor selected by the clustering factor selection step to divide the cluster into clusters;
A factor determining step of determining an input factor and an output factor of the model for each cluster using the factors included in the cluster,
A modeling step of modeling the input / output relationship for each cluster using the time series data corresponding to the input factor and the output factor included in the cluster,
A model accuracy evaluation step of evaluating the accuracy of an integrated model integrating a plurality of models constructed for each cluster,
A cluster correction step of correcting the cluster so that the evaluation of the accuracy of the integrated model satisfies a predetermined condition;
With
Until the evaluation of the accuracy of the integrated model satisfies a predetermined condition, a cluster correction step, a factor determination step, a modeling step, and a model accuracy evaluation step are sequentially performed to construct a model.

Further, the adaptive prediction model building method according to the invention described in claim 5 is as follows.
The adaptive prediction model building method according to claim 4,
The clustering factor selection step calculates a value representing the center of each factor for each of the preliminary clusters generated by the preliminary clustering, calculates the importance of the clustering factor based on the degree of dispersion of the values, and calculates the clustering factor. It is characterized by selecting.

Further, the adaptive prediction model building method according to the invention of claim 6 includes:
The adaptive prediction model building method according to claim 4,
The clustering factor selection step calculates the importance of the clustering factor according to the degree of overlap of the range from the minimum value to the maximum value of each factor for each preliminary cluster generated by preliminary clustering, and selects the clustering factor. It is characterized by.

Further, an adaptive prediction model building method according to the invention of claim 7 comprises:
The adaptive prediction model building method according to claim 4,
The preliminary clustering step, the clustering factor selection step, and, instead of the main clustering step,
A feature is that a plurality of time-series data groups are created by subdividing time-series data that can be used for construction of a prediction model in advance, and the model configuration is determined by using this as an initial division.

The adaptive prediction model construction system according to the invention described in claim 8 is:
In an adaptive prediction model building system such as a computer that adaptively builds a prediction model for predicting future time series data based on past time series data,
Prediction model construction means for constructing a prediction model using past time-series data,
A prediction execution unit that executes prediction using the prediction model and obtains a prediction value;
Prediction value correction means for calculating a correction coefficient or a correction amount for correcting the prediction value based on a past prediction error or a model error,
Prediction error, model error, based on the correction coefficient or the correction amount, whether the model needs to be updated, or whether the model configuration needs to be updated, and a prediction model update determining means for determining
When it is determined that model update and model configuration update are unnecessary, predicted value correction executing means for correcting the predicted value,
Predictive model updating means for updating the predictive model using newly accumulated time-series data when it is determined that model update is necessary;
Predictive model configuration updating means for updating the predictive model configuration using newly accumulated time-series data when it is determined that the model configuration needs to be updated;
It is characterized by having.

Further, the adaptive prediction model construction system according to the invention of claim 9 comprises:
The adaptive prediction model construction system according to claim 8,
When the prediction error, the model error, the correction coefficient or the correction amount exceeds the allowable value for model update determination, it is determined that the prediction model needs to be updated, and the tendency / characteristics of the time series data to be predicted change. Alarm output means for outputting an alarm for notifying that
It is characterized by having.

The adaptive prediction model construction system according to the invention according to claim 10 is:
The adaptive prediction model construction system according to claim 8,
When the prediction error, the model error, the correction coefficient, or the correction amount exceeds the allowable value for model configuration update determination, it is determined that the prediction model configuration needs to be updated, and the trend / characteristics of the time series data to be predicted changes. Alarm output means for outputting an alarm for notifying that
It is characterized by having.

An adaptive prediction model construction system according to the invention described in claim 11 is:
In the adaptive prediction model construction system according to any one of claims 8 to 10,
The prediction model construction means and the prediction model configuration updating means,
Means for analyzing time series data of a plurality of factors, finding an input factor having an input / output relationship with a certain output factor from among the plurality of factors, and modeling the input / output relationship,
Preliminary clustering means for performing clustering using all of the time-series data to be used and dividing the cluster into preliminary clusters;
Clustering factor selecting means for calculating the importance of the clustering factor for each preliminary cluster and selecting the clustering factor;
A clustering unit that performs clustering using the time-series data related to the clustering factor selected by the clustering factor selection unit and divides the cluster into clusters;
A factor determining means for determining an input factor and an output factor of the model for each cluster using the factors included in the cluster,
Modeling means for modeling the input / output relationship for each cluster using time series data corresponding to the input factor and the output factor, which are included in the cluster,
A model accuracy evaluation means for evaluating the accuracy of an integrated model obtained by integrating a plurality of models constructed for each cluster,
Cluster correction means for correcting the cluster so that the evaluation of the accuracy of the integrated model satisfies a predetermined condition;
And a model is constructed by sequentially operating the cluster correction unit, the factor determination unit, the modeling unit, and the model accuracy evaluation unit until the evaluation of the accuracy of the integrated model satisfies a predetermined condition.

An adaptive prediction model construction system according to the invention of claim 12 is:
The adaptive prediction model construction system according to claim 11,
The clustering factor selection means calculates a value representing the center of each factor for each of the preliminary clusters generated by the preliminary clustering, calculates the importance of the clustering factor based on the degree of variation of the value, and calculates the clustering factor. It is a means for selecting.

An adaptive prediction model construction system according to the invention of claim 13 is:
The adaptive prediction model construction system according to claim 11,
The clustering factor selection means calculates the importance of the clustering factor according to the degree of overlap of the range from the minimum value to the maximum value of each factor for each preliminary cluster generated by preliminary clustering, and selects the clustering factor. There is a feature.

The adaptive prediction model construction system according to the invention according to claim 14 is:
The adaptive prediction model construction system according to claim 11,
In place of the preliminary clustering unit, the clustering factor selection unit, and the main clustering unit,
A feature is that a plurality of time-series data groups are created by subdividing time-series data that can be used for construction of a prediction model in advance, and the model configuration is determined by using this as an initial division.

According to the present invention as described above, it is possible to provide an adaptive prediction model construction method or an adaptive prediction model construction system that enables an accurate model to be efficiently constructed using time-series data.

Hereinafter, the best mode of the present invention will be described. First, an adaptive prediction model construction method will be described with reference to the drawings.
FIG. 1 is a flowchart for explaining a process of constructing an adaptive prediction model, FIG. 2 is a flowchart of a prediction model update step, FIG. 3 is a flowchart of a prediction model configuration update step, and FIG. 4 is a process of a prediction model construction step / prediction model configuration update step. It is a flowchart explaining.

The adaptive prediction model construction method of the present embodiment analyzes time series data of a plurality of factors, finds an input factor having an input / output relationship with a certain output factor from the plurality of factors, and models the input / output relationship. It is. Hereinafter, an adaptive prediction model construction method will be described with reference to FIG.

(1) Prediction model construction step Step S1 is a prediction model construction step of constructing a prediction model using past time-series data. Specifically, necessary time-series data is input and stored, and a prediction model is constructed. The prediction model constructed here is used as an initial prediction model. The method of constructing the initial prediction model will be described later in detail with reference to FIGS. 4 to 8, and the entire process will be described first.

(2) Prediction execution step Step S2 is a prediction model execution step of executing prediction using a prediction model and obtaining a predicted value.
With respect to the prediction model constructed in the prediction model construction step (1), time series data for prediction is input, and a prediction calculation of the time series data is executed to obtain a predicted value.

(3) Prediction value correction step Step S3 is a prediction value correction step of correcting the prediction value based on a past prediction error or a model error.
In order to correct the predicted value calculated in step S2, a past prediction error (model error) is calculated.

Here, when uncertain information such as a temperature forecast value is not used as input information to the prediction model, a normal prediction error may be calculated.
On the other hand, when uncertain information is used as input information to the prediction model, a model error is calculated using the actual value of the information (actual temperature value in the case of temperature). This is to eliminate the influence of the uncertain information input to the prediction model when correcting the prediction value.

A correction coefficient and a correction amount are calculated based on the calculated prediction error.
If it is a correction coefficient, it can be calculated, for example, as follows: correction coefficient = 100 / (100 + error%).
The prediction value is corrected using the correction coefficient and the correction amount.

(4) Prediction model update determination step Step S4 is a prediction model update determination step of determining whether the model needs to be updated or the model configuration needs to be updated based on the past prediction error or the model error.
The above (2) and (3) constitute a loop executed every time the prediction timing of the time-series data is concerned. In the loop, it is determined whether or not the model needs to be updated. This judgment determines whether only the prediction model needs to be updated or whether the update including the prediction model configuration is necessary.

The judgment can be made using a prediction error (model error), a correction coefficient, a correction amount, and the like. For example, in the case of a prediction error, when a prediction error equal to or more than a set allowable value continues for a certain period of time, or when an average of the closest prediction errors is equal to or more than a threshold, when constructing an initial prediction model, Since it is determined that the relationship between the used data and the current data has changed, it can be determined that the prediction model needs to be updated. The same applies to the correction coefficient and the correction amount.

Here, if the model update determination allowable value used for model update determination and the model configuration update determination allowable value used for model configuration update determination are set to the same value, not only the model but also the model configuration is always updated.
In addition, if the allowable value for model update determination used for model update determination is set smaller than the allowable value for model configuration update determination used for model configuration update determination, only an update of the model is performed if the error is within a certain fixed value. When the error is equal to or more than a certain value, the method of updating the model can be adjusted according to the magnitude of the error so that the error including the model configuration is updated.

(5) Predicted Value Correction Execution Step Step S5 is a predicted value correction execution step of correcting the predicted value when it is determined that the model update and the model configuration update are unnecessary. The predicted value is corrected and output.

(6) Prediction model update step In step S6, when it is determined that the model update is necessary in the prediction model update determination step, the prediction model is updated using the newly accumulated time series data for updating the prediction model. This is an update step.
The prediction model update step includes an alarm output step and an update execution step, as shown in FIG.

In step S61, when the prediction error, the model error, the correction coefficient or the correction amount exceeds the allowable value for model update determination, it is determined that the prediction model needs to be updated, and the trend / characteristics of the time series data to be predicted are determined. Is an alarm output step for performing an alarm output for notifying that is changing.
In the above-described prediction model update determination, an alarm is output to the operator by determining that the trend of the time-series data to be predicted has changed, that the prediction model needs to be updated, and the like.
Step S62 is an update execution step of updating the prediction model using the newly accumulated time-series data.

(7) Prediction model configuration update step In step S7, when it is determined in the prediction model update determination step that the model configuration needs to be updated, the prediction model configuration is updated using newly accumulated time-series data. This is an update step.
The prediction model configuration update step includes an alarm output step and a configuration update execution step, as shown in detail in FIG.

In step S71, when the prediction error, the model error, the correction coefficient, or the correction amount exceeds the allowable value for model configuration update determination, it is determined that the prediction model configuration needs to be updated, and the trend of the time series data to be predicted is determined. An alarm output step for outputting an alarm for notifying that the characteristic has changed.
In the above-described model configuration update determination, an alarm is output to the operator by determining that the trend of the time-series data to be predicted has changed or that the prediction model configuration needs to be updated.
Step S72 is a configuration update execution step of updating the prediction model configuration using the newly accumulated time-series data.

(8) Prediction target period end determination step Step S8 is a prediction target period end determination step.
If it is determined that the prediction target period has not ended, the process jumps to the beginning of step S2 and repeats the processing of steps S2 to S7. On the other hand, if it is determined that the prediction target period has ended, the process ends.

In the present embodiment, the prediction model construction step and the prediction model configuration update step follow the flow as shown in FIG. 4 in detail. Such a process will be described with reference to the drawings.
In the prediction model construction step / prediction model configuration update step, cluster analysis is adopted. A method of generating a plurality of clusters composed of time-series data having relevance by this cluster analysis, constructing a model for each cluster, and finally constructing an integrated model integrating the plurality of models is adopted.

(1) Initial setting step Step S10 is an initial setting step. Specifically, step S10 is for selecting whether or not to select a variable for clustering, setting a clustering method, setting a modeling method, setting a cluster division correction method, and setting other parameters. Step.
Whether or not to select a variable for clustering means that the user or developer who builds the model selects whether or not to explicitly determine the factor to be used for clustering from all the factors that can be used to build the model. That's what it means.

The setting of the clustering method selects, for example, a method using a hierarchical method or a method using a non-hierarchical method.
The setting of the modeling method selects a prediction model such as a neural network or a multiple regression equation.
The setting of the cluster division correction method determines whether to use a method based on a correlation coefficient described later or an analysis based on error accuracy.
In the parameter setting, necessary parameters other than the above (for example, the number of divided data spaces in the case of top-down correction described later) and other specific numerical values are input.

(2) Preliminary clustering step Step S11 is a preliminary clustering step of performing clustering using all of the time-series data to be used and dividing into clusters.
Preliminary clustering is performed in order to obtain a preliminary cluster for obtaining the center and cluster range of the cluster used in the clustering factor selection process performed in the next step S12 (3) Clustering factor selection step.

(3) Clustering Factor Selection Step Step S12 is a clustering factor selection step of calculating the importance of the clustering factor for each backup cluster and selecting a clustering factor.
The degree of separation is calculated from the clusters obtained as a result of the preliminary clustering in step S11 of the above (2), and a clustering factor is selected using this.

The degree of separation calculated in the clustering factor selection step includes (A) a degree of separation using a cluster center and (B) a degree of separation using a cluster range.
Hereinafter, both will be described sequentially.

(A) Selection of Factor for Clustering Based on Degree of Separation Using Cluster Center Here, the degree of separation using the center of cluster will be described. In general, clustering is performed based on a multidimensional Euclidean distance instead of a specific factor, and thus the clustering result differs depending on the factor used and the distribution state of data.

Here, it is considered to find an important factor for cluster division by analyzing the distribution of each factor for each cluster. Specifically, the degree of separation using the center value of the cluster is defined for each factor, and a factor having a high degree of separation is selected as a factor for clustering (or a factor having a low degree of separation is excluded from the factors for clustering).
The degree of separation using the center of a cluster is defined as the variation of the center value of each cluster for each factor as shown in Equation 1 of the following equation, for example.

A specific example will be described. FIG. 5 is an explanatory diagram of the degree of separation using the center of the cluster. It is assumed that clustering is performed using two factors of the factor χ _i and the factor _{ｊ j} , and as a result, the cluster is divided into three clusters C1, C2, and C3. Here, which of the factors _{ｉ i} and χ _j is important for clustering is determined using the center of each cluster. Specifically, it can be said that the larger the width of the value of the center avg of the factor 大 き い, the more the cluster is separated. In the example of FIG. 5, the factor _{ｉ i} has a higher degree of separation than the factor _{ｊ j} , and can be determined to be an important factor for clustering.

(B) Degree of Separation Using Cluster Range Here, the degree of separation using the cluster range will be described.
The degree of separation using the cluster range is a method in which the degree of cluster overlap is considered, and a factor in which the range of the divided clusters has little overlap is considered to be a large degree of separation, that is, a factor important for clustering.

A specific example will be described. FIG. 6 is an explanatory diagram of the degree of separation using the range of the cluster. It is assumed that clustering is performed using two factors of the factor χ _i and the factor _{ｊ j} , and as a result, the cluster is divided into three clusters C1, C2, and C3. Here, which of the factors _{ｉ i} and χ _j is important for clustering is determined using the range of each cluster.

First, it is determined whether or not the cluster C1 has a range overlapping with the clusters C2 and C3 for each factor. A description will be given using the factor _ｉi as an example. First, the entire range of values factor chi _i, obtaining the _{A i.} In the illustrated example, represent the range _{A i} of the total value factor chi _i, as a range between χ _i ^min (C1∪C2∪C3) and χ _i ^max (C1∪C2∪C3).

Next, a product set is obtained for the clusters C1 and C2, the clusters C1 and C3, and the clusters C2 and C3 (the combination of the two clusters is performed for all clusters). The union of the product set obtained here is determined, the minimum value _{ｉ i} ^min and the maximum value _{ｉ i} ^max of the factor χ _i in the union are determined, and the overlapping range L = maximum value−minimum value is determined. In the example of FIG. 6, since there is no intersection of clusters C1 and C3 and clusters C2 and C3, the intersection of clusters C1 and C2, that is, the range where C1 and C2 overlap is L. The degree of separation of clusters in the factor _ｉi is defined, for example, by the following equation.

Similar processing is performed for the factor χ _j to determine the degree of separation DL _j . This DL indicates the ratio of the range of the overall value of each factor to the range in which clusters do not overlap. Therefore, a factor with a smaller DL indicates that the clusters overlap less, that is, that the clusters are clearly separated, and can be determined to be an important factor for clustering.

(C) Selection of Factor for Clustering The factor for clustering is selected using the degree of separation obtained by any of the above (A) and (B). Specifically, a factor whose degree of separation is smaller than a preset threshold is not a factor effective for clustering, and is therefore excluded from the clustering factors. As a result, the remaining factors are used as clustering factors, and used in the next step of the main clustering. By selecting this clustering factor, more appropriate clustering can be performed.

(4) Main Clustering Step Step S13 is a main clustering step of performing clustering using the time-series data related to the clustering factor selected in the clustering factor selecting step to divide the cluster into clusters.
By this clustering, a multidimensional data space represented by some or all of the factors is divided into a plurality of clusters.

The cluster division will be described with reference to the drawings. FIG. 7 is an explanatory diagram illustrating generation of a cluster.
It is assumed that the data space is a three-dimensional data space and is divided as shown in FIG. In this division example, each cluster includes data as shown in the table of FIG. In the table of FIG. 7B, “□ (i)”, “× (i)”, “△ (i)”, and “○ (i)” represent some values other than 0 (zero). Shall be.
In cluster 1 and cluster 2, the space is such that the values of factors 1 and 2 are other than 0 and the value of factor 3 is 0.
Cluster 3 is a space in which the values of factors 1 and 3 are other than 0, and the value of factor 2 is 0.
Cluster 4 is a space in which the values of factors 1, 2, and 3 are other than 0.
Clusters 1, 2, and 3 show a case where a space whose value is 0 is generated by one of the factors.
Clusters 1 and 2 are common when the value of factor 3 is 0, but show a case where another cluster is generated.

(5) Factor determination step Step S14 is a step of determining an input factor and an output factor of the model for each cluster using the factors included in the cluster.
By evaluating the cluster generation result and the information on the cluster division process, an input factor used in the modeling of (6) described later is determined.

For example, in the cluster shown in FIG. 7, when the factor 1 is selected as the output factor, the input factor is the factor 2 in the clusters 1 and 2, the input factor 3 in the cluster 3, and the input factor is the factor in the cluster 4. It is determined to be two or three. In these clusters 1 to 3, the input factors are different even for the same output factor, and the clusters are generated such that the input factors are different when various conditions are different. In this way, a cluster includes all factors or a space including only specific factors is constructed. Further, in a cluster, a plurality of spaces may be created even if the factors are the same. This is a phenomenon that can occur in cluster analysis.

(6) Modeling Step Step S15 is a modeling step of modeling an input / output relationship for each cluster using time-series data corresponding to an input factor and an output factor, which are time-series data included in the cluster. It is.
For modeling, typical methods are selected from the use of past cases, a multiple regression equation, a neural network, and the like.
At this time, the input factors of the model are determined from the factors included in the cluster. In determining this input factor, a method of selecting the model developer or user based on empirical judgment, a method of further limiting the input factor by using an information amount criterion such as AIC (Akaike Information Criterion), and the like. Is applicable.

In using past cases, for example, there is a method in which the average value of the time-series data included in the obtained cluster is used as the output value of the model.
In the use of the multiple regression equation, a multiple regression analysis is performed using the time series data included in the obtained cluster to construct a multiple regression equation.
Similarly, in the case of using a neural network, learning of the neural network is performed using the time-series data included in the obtained cluster to form a model.

(7) Model accuracy evaluation step Step S16 is a model accuracy evaluation step of evaluating the accuracy of an integrated model obtained by integrating a plurality of models constructed for each cluster.
The accuracy evaluation of the integrated model is performed by the following equation.

If the output error EE of Equation 3 is less than a predetermined value, it is determined that model rebuilding is unnecessary. If the output error EE is more than a predetermined value, it is determined that model rebuilding is necessary. .
Here, for each model i, it is possible to consider the degree of importance of accuracy evaluation for each model using the weighting coefficient wi. If the adjustment of the weight coefficient wi is set to a binary value of 0 and 1, it is possible to perform an accuracy evaluation in which only an evaluation for a specific model is performed.

Note that the multivariate data used for model accuracy evaluation (error evaluation) includes a method of evaluating using a modeling error that uses the multivariate data as it is when performing modeling, and a method of evaluating all time-series data when modeling. There is a method that does not use a part of the time-series data and saves it for model accuracy evaluation.
In the former method, it is possible to evaluate how much the time-series data given for modeling has been modeled.
In the latter method, the generalization ability of the model can be evaluated.

(8) Cluster Correction Step Step S17 is a cluster correction step of correcting a cluster so that the evaluation of the accuracy of the integrated model satisfies a predetermined condition. As a method for correcting such clusters, for example, a correction performed in a bottom-up manner (hereinafter, particularly referred to as a bottom-up correction in the present specification and described below), and a correction performed in a top-down manner (hereinafter, referred to as “bottom-up correction”). In this specification, bottom-up correction and its name are particularly defined and described below.

First, the bottom-up correction will be described. The bottom-up correction calculates an index indicating a correlation of an input / output relationship, such as a correlation coefficient, for the time-series data included in the cluster. Table 1 is a table showing the correlation coefficient.

As shown in Table 1, it is assumed that the correlation coefficients of the input and output data in the clusters C1, C2, and C3 are 0.6, 0.7, and 0.5, respectively.
Next, a combination of each cluster is generated, and a new cluster obtained by the combination is set as a temporary cluster. In Table 1, three types of combinations are generated as temporary cluster combinations.
Subsequently, a correlation coefficient is calculated for the time-series data included in the temporary cluster. In the example of Table 1, four types of clusters and their respective correlation coefficients were obtained together with the initial cluster.

Here, it is determined which of the four data divisions is to be selected based on the correlation coefficient. As a determination method, for example, there is a method of selecting a division including a cluster having the largest correlation coefficient. In the case of Table 1, the temporary cluster 3 is selected because it includes a cluster having a correlation coefficient of 0.9.
There is also a method of selecting a division having the highest average correlation coefficient. In the case of Table 1, the temporary cluster 1 is selected because the average of the correlation coefficients is 0.7.
Furthermore, it is conceivable to attach importance to the correlation coefficient of a cluster including time-series data related to a specific factor.

Such a cluster correction step is performed until all clusters are finally merged and coincide with the original data space (n-dimensional space which is a multidimensional space and is determined by n factors). From the obtained clusters, a cluster having the highest overall accuracy of the integrated model and its model are selected. This correction is defined as bottom-up correction in this specification.

Next, the top-down correction will be described.
Top-down correction is a method in which an error between a model output obtained by inputting time-series data for an input factor of a model determined for each cluster and time-series data related to an output factor is equal to or more than a predetermined threshold. This is the step of correcting the cluster by dividing it into a new data space.

Top-down correction uses the error in each model calculated in the accuracy evaluation of the entire model. The error calculated here represents the modeling accuracy of the current model. In other words, it indicates that a model having a large error cannot be appropriately modeled. Therefore, the time-series data in which the error is equal to or greater than a certain threshold is separated as a new cluster.

Hereinafter, a specific example of the top-down correction will be described with reference to the drawings.
FIG. 8 is an explanatory diagram illustrating correction of a cluster.
Although the data space is actually a multidimensional data space, FIG. 8 illustrates the data space as a two-dimensional data space having only two factors, factor 1 and factor 2, for ease of explanation.
Here, by executing each phase of step S10 to step S15, the entire data space is divided into three clusters, cluster 1, cluster 2, and cluster 3. And the cluster 3 include five time-series data.

The models 1, 2, and 3 representing the input / output relationship of the time-series data for each cluster are constituted by a neural network, a multiple regression equation, and the like. In the model accuracy evaluation step described above, individual data in the time-series data is input to the model 1, model 2, and model 3 corresponding to each cluster, and a modeling error is calculated. Here, it is assumed that the multivariate data in a case where the modeling error is large (a case where the modeling error is larger than a separately set threshold) are circles (d1 to d7) in FIG.

時 The existence of such time-series data indicates that modeling cannot be performed properly with the current model for each cluster. Therefore, the cluster is corrected. As a method of correction, a method of (a) separating a new cluster from an original cluster and a method of (b) separating all time-series data separated from each original cluster into one new cluster are considered. Can be

In the method (a), a cluster is generated that includes d1, d2 in a new cluster 1 ′, d3, d4, d5 in a new cluster 2 ′, and d6, d7 in a new cluster 3 ′. It is. In this case, the original cluster 1 from which d1 and d2 have been removed, the original cluster 2 from which d3, d4, and d5 have been removed, and the original cluster 3 from which d6 and d7 have been removed remain.
In the method (b), a new cluster 4 having d1 to d7 as one is generated. In this case, the original cluster 1 from which d1 and d2 have been removed, the original cluster 2 from which d3, d4, and d5 have been removed, and the original cluster 3 from which d6 and d7 have been removed remain.

(4) In such a cluster correction step, the cluster is finally divided into the set number, and a cluster having the highest overall accuracy of the integrated model and the model are selected from the clusters obtained in the process. Such a correction is defined as a top-down correction in this specification.

(6) Model Reconstruction Step In steps S14 to S17, the cluster correction step, the factor determination step, the modeling step, and the accuracy evaluation step are repeatedly performed until the evaluation of the accuracy of the integrated model satisfies a predetermined condition. This is a model rebuilding step for building a model.
Evaluation of the accuracy of the integrated model, that is, when the output error EE of Equation 3 does not fall below a predetermined value, the cluster is changed until the EE falls below a predetermined value, and the factors and times included in the cluster are changed. Model construction based on the series data is repeated.

According to the present embodiment described above, a data space constituted by factors of time-series data is divided to generate clusters, a cluster including an output factor and an input factor related to the output factor is extracted, and an output factor is extracted. Since a model based on these output factors and input factors is constructed based on the time series related to the output factors and the output factors, a highly accurate model can be constructed.

Instead of performing steps S10 to S13, a plurality of time-series data groups are created by subdividing the time-series data available for the construction of the prediction model in advance, and this is used as an initial division to determine the model configuration. It may be performed. In this case, empirical knowledge by a specialist is involved in place of steps S10 to S13 shown in FIG. 4, but the optimum is determined later by a factor determination step, a modeling step, a modeling accuracy evaluation step, and a cluster correction step. Model will be constructed. Such a form can also be adopted.

Next, an adaptive prediction model construction system will be described. FIG. 10 is a system configuration diagram of the adaptive prediction model construction system.
As shown in FIG. 10, the adaptive prediction model construction system includes a data input screen 1, a data input unit 2, a data storage unit 3, a database 4, a data output unit (display device) 5, a prediction model construction unit 6, and a prediction model update. Means 7, prediction model configuration update judging means 8, prediction model configuration updating means 9, alarm output means 10, prediction model 11, prediction execution means 12, prediction value correction execution means 13, and prediction value correction means 14.
Further, a network 100 such as the Internet, an intranet, and a LAN is connected to the data input unit 2.
The adaptive prediction model construction system corresponds to a computer (computer) on which these means function.

(1) Prediction model construction means 6
This is a means for constructing a prediction model using past time-series data, and has a function corresponding to step S1. Specifically, it is a means for inputting / accumulating necessary time-series data and constructing a prediction model. Necessary time-series data includes a data input screen of a computer, data stored in another computer in a LAN to which the system is connected, and data stored in another computer via the Internet. The data is input into the system through the data input unit 2 and stored in the database 4 by the data storage unit 3. The prediction model 11 is constructed using the data in the database 4 thus accumulated. The constructed prediction model 11 is used as an initial prediction model.

(2) Prediction executing means 12
This is a means for executing prediction using the prediction model 11 and obtaining a predicted value, and has a function corresponding to step S2.
To the prediction model 11 constructed by the prediction model construction means 6 of (1), time series data for prediction is inputted, and a prediction calculation of the time series data is executed to obtain a predicted value.

(3) Prediction value correction means 14
This is a means for correcting the predicted value based on a past prediction error or a model error, and has a function corresponding to step S3.
In order to correct the predicted value calculated by the prediction executing means 12, a past prediction error (model error) is calculated.

(4) Prediction model update determination means 8
This is a means for determining whether the model needs to be updated or the model configuration needs to be updated based on the past prediction error or model error, and has a function corresponding to step S4.
The above (2) and (3) constitute a loop executed every time the prediction timing of the time-series data is concerned. In the loop, it is determined whether or not the model needs to be updated. This judgment determines whether only the prediction model needs to be updated or whether the update including the prediction model configuration is necessary.

(5) Predicted value correction executing means 13
The predicted value correction execution unit 13 corrects the predicted value when it is determined that the model update and the model configuration update are unnecessary, and has a function corresponding to step S5. The predicted value is corrected and output. The final prediction value obtained here and the data used for the prediction are stored in the database 4 and output through the data output means (display device) 5.

(6) Prediction model updating means 7
When the prediction model update determining means 8 determines that the model needs to be updated, the prediction model updating means updates the prediction model using the newly accumulated time-series data for updating the prediction model, and corresponds to step S6. .
The prediction model update means 7 includes an alarm output means and an update execution means in detail.

The alarm output means determines that the prediction model needs to be updated when the prediction error, the model error, the correction coefficient or the correction amount exceeds the allowable value for model update determination, and determines the trend / characteristics of the time series data to be predicted. Is a means for outputting an alarm for notifying that is changing, and has a function corresponding to step S61 in FIG.
In the above-described prediction model update determination, an alarm is output to the operator by determining that the trend of the time-series data to be predicted has changed, that the prediction model needs to be updated, and the like.
The update execution unit is a unit that updates the prediction model using the newly accumulated time-series data, and has a function corresponding to step S62.
The prediction model updating means 7 functions to update the prediction model 11.

(7) Prediction model configuration updating means 9
The predictive model configuration updating means is means for updating the predictive model configuration using newly accumulated time series data when the predictive model update determining means determines that the model configuration needs to be updated, and corresponds to step S7. It has a function to do.
Specifically, the prediction model configuration updating unit includes an alarm output unit and a configuration update executing unit.

The alarm output means determines that it is necessary to update the prediction model configuration when the prediction error, the model error, the correction coefficient or the correction amount exceeds the model configuration update determination allowable value, and determines the time series data of the prediction target. This is a means for outputting an alarm for notifying that the tendency / characteristic has changed, and has a function corresponding to step S71.
In the above-described model configuration update determination, an alarm is output to the operator by determining that the trend of the time-series data to be predicted has changed or that the prediction model configuration needs to be updated.
The configuration update execution unit is a unit that updates the prediction model configuration using the newly accumulated time-series data, and has a function corresponding to step S72.
The prediction model configuration updating means 9 functions to update the prediction model 11.

The data output means (display device) 5 outputs alarm information from the alarm output means 10 in addition to the predicted value obtained by the prediction execution means and the data used for the prediction.
The adaptive prediction model construction system of the present embodiment has the above configuration. The adaptive prediction model construction system includes, in addition to the system shown in FIG. 10, means corresponding to each step shown in FIGS. 2 to 4, and is a system that constructs an adaptive prediction model by these means. Is also good.

Next, embodiments of the present invention will be described using specific examples. It is assumed that the power demand on the next day and one day in a certain region to be predicted is predicted. Various methods have been devised for power demand forecasting. As an example, here, a forecast model is constructed using time-series data of 2000 (previous year), and the power demand of 2001 (next year) is calculated. Think about predicting.

(1) Prediction model construction means 6
Build a prediction model for predicting time series data. Here, an example using the model construction system of the invention according to claim 14 will be described.
The power demand (prediction target) in 2000 is assumed as time-series data necessary for constructing a prediction model, and the input factors are, for example, a maximum temperature and a minimum temperature. A neural network is used as a prediction model.

A specific description will be given with reference to FIG. 10 illustrating an example of a system configuration according to the embodiment of the present invention. The power demand of each customer is separately measured. For the power demand to be predicted in this embodiment, the sum of the power demand of each customer is calculated, and the power demand is calculated by the network 100 such as the Internet, an intranet, or a LAN. It is stored in a database (not shown) in another computer connected to the system of the invention. The maximum and minimum temperatures used as prediction inputs are used for past (2000 in this example) maximum and minimum temperature actual values used in the prediction model construction, and for prediction of the next day. The highest temperature forecast value and the lowest temperature forecast value of the next day are obtained from the weather company, and these data are accumulated in the database 4. In general, power demand data and weather data can be obtained by adopting such a form. However, it is of course possible to input data from the data input screen 1 to a computer instead of the Internet or the like.

Here, learning is performed by giving these time series data to input factors of the neural network. That is, learning is performed so that the maximum temperature and the minimum temperature are input and the power demand is output, and the initial prediction model 11 is constructed.
Here, for the sake of simplicity, the input of the neural network is set to the highest temperature and the lowest temperature. However, in order to improve the prediction accuracy, a factor considered to affect the power demand may be used. For example, the actual power demand value of the previous day, the predicted humidity of the day, the predicted weather, the predicted solar radiation, flag data indicating whether the predicted date is a weekday, a Saturday, or a holiday can be considered.

In this case, as a configuration of the initial prediction model 11, it is considered that the power demand has different characteristics for each season.
The seasonal characteristics of the electric power demand have a positive correlation that the demand for cooling is high in summer and that the higher the temperature, the higher the electric power. Because of the high demand for heating in winter, there is a negative correlation between low temperatures and high power. In addition, there is little correlation between temperature and electricity because there is almost no cooling and heating demand in spring and autumn.

Therefore, the prediction model 11 is divided for each season, and the prediction model 1 (spring) for April 1 to June 30, the prediction model 2 (summer) for July 1 to August 31, and the September 1 to November 30 is a prediction model 3 (autumn), and December 1 to March 31 is a prediction model 4 (winter).
Learning is performed on these prediction models 1 to 4 using the maximum temperature and the minimum temperature in each period.

(2) Prediction executing means 12
The time series data for prediction is input, and the prediction calculation of the time series data is executed using the prediction model 11 constructed in (1). In this example, the data input for prediction are the maximum temperature and the minimum temperature, but the maximum temperature and the minimum temperature of the next day are not known. You will enter a temperature forecast. By inputting the above-mentioned time-series data into the prediction model of the neural network, a predicted value of the power demand on the next day can be obtained.

(3) Prediction value correction means 14
Correction is performed on the predicted value calculated in (2). A past prediction error (model error) is calculated to perform the correction. Here, since the input information to the prediction model 11 is a temperature forecast value, that is, uncertain information, the model error is calculated using the actual value of the information (actual temperature value in the case of temperature).

This is to exclude the influence of uncertain information input to the prediction model 11 when correcting the prediction value. That is, after the next day (after the highest temperature, the lowest temperature, and the power demand on the previous day are found), actual values are input to the prediction model 11 instead of the forecast values, and prediction is performed using the actual values. The estimated power demand value and its error (model error) are calculated.

補正 Calculate the correction coefficient and correction amount based on the calculated prediction error or model error. The purpose here is to correct the error in the prediction model at the time of the prediction so that the error is eliminated in a certain direction such that the error is high and low. Therefore, it is necessary to numerically express the degree of deviation between the prediction error and the model error in the closest period.

Specifically, the average error (E) is calculated for the model errors excluding the maximum error and the minimum error from the model errors for the past N days. In the example of Table 2 above, when N is 4, when the error is 7/4 to 7/7, the error for 4 days of -1.2%, 3.2%, 3.0%, and 3.3% is the minimum value. Calculate the average of 3.2% and 3.0% excluding a certain -1.2% and the maximum value of 3.3% to obtain 3.1%.

Next, a correction coefficient is calculated using this error. The correction coefficient is, for example, as follows.

で は In the case of this example, since the average of the error is 3.1%, the correction coefficient can be calculated as in the following equation.

予 測 The predicted value can be corrected by multiplying the predicted value by the correction coefficient. Without correction, an error of about 3% appears on the plus side, but this error can be reduced by multiplying by a correction coefficient of 0.9699.

(4) Prediction model update determination means 8
The model update determination can be performed using a prediction error (model error), a correction coefficient, a correction amount, and the like. For example, in the case of a prediction error, when a prediction error equal to or more than a set allowable value (model update determination allowable value) continues for a certain period, or when an average of the closest prediction errors is equal to or more than a threshold value, Since it is determined that the relationship between the data used to construct the initial prediction model and the current data has changed, it can be determined that the prediction model needs to be updated. The same applies to the case where a correction coefficient and a correction amount are used.

For example, consider a case where a model update decision is made using a prediction error. As the prediction error, as shown in the above (3), the average (E) of the errors for N-2 days excluding the minimum value and the maximum value of the model error from the model error for the nearest N days is used. Here, assuming that the threshold used for the model update determination is E1 and the threshold used for the model configuration update determination is E2, the relationship between the error E and E1 and E2 is as follows.

IF E ≤ E1 THEN Executes correction of predicted value IF E1 <E ≤ E2 Executes update of THEN model IF E2 <E THEN Executes update of model configuration

Assuming that E1 is set to 4% and E2 is set to 5%, since E = 3.1% in the above-described example, the model and the model configuration are not updated, and the prediction value is corrected. .

(5) Predicted value correction executing means 13
Correct the predicted value. The final predicted value obtained here and the data used for the prediction are stored in the database 4 and output through the data output means (display device) 5 which is a computer display. In the case of this embodiment, the final power demand predicted value and the data (highest temperature, lowest temperature) used for the prediction are output to the data output means (display device) 5.

(6) Prediction model updating means 7
In the model update determination of (4), when it is determined that the model needs to be updated, the prediction model is updated using the newly stored data for updating the prediction model. In the above example, E1 is 4% and E2 is 5%. However, for example, if the set value of E1 is 3% and E2 is 5%, the average error E is 3.1%, which is larger than E1. , A model update is performed. In the model update, the prediction model for summer is re-learned using the data from July 1 to July 7, which is newly accumulated data in the summer period, to update the model. Here, the model may be updated using only the newly stored data, or may be used together with the 2000 summer learning data.

In this case, in the above-described model update determination, an alarm is notified to the operator that the trend of the prediction target time series has changed, that the prediction model needs to be updated, and that the prediction model configuration needs to be updated. I do.

(7) Prediction model configuration updating means 9
In the model update determination of (4), when it is determined that the model configuration needs to be updated, the prediction model configuration is updated using the data for updating the prediction model configuration.
This updating process is performed by the updating process of the model configuration shown in FIG.
In the above-described example, E1 and E2 used to determine the model configuration are determined to be 4% and 5%. However, for example, if the set value of E1 is 2% and the set value of E2 is 3%, the average error E is Since 3.1% is larger than E1 and E2, the model configuration is updated.

In the case of this example, the prediction error has increased since July 2001, and it has been determined that correction of the prediction value and updating of the model using newly accumulated data cannot be improved. The model configuration is updated, that is, the seasonal segments are reconfigured.

In this case, in the above-described model update determination, the operator is notified of an alarm indicating that the trend of the prediction target time series is changing and that the prediction model configuration needs to be updated.
More specifically, the model error for the past N days, the maximum error, the minimum error, and the maximum error in the modeling error for the N days, the average error (E) excluding the minimum error, and the correction coefficient calculated using the average error , E1 used for model update determination, E2 used for model configuration update determination, and information on "whether correction has been performed", "whether model update is required", or "model configuration update is required". In the example of Table 2 above, the error of 7/4 to 7/7 is -1.2%, 3.2%, 3.0%, and 3.3% for four days, the minimum error is -1.2%, and the maximum error is 3.3%. The average error of 3.1%, the value of E1 is 2%, the value of E2 is 3%, and the fact that "model configuration needs to be updated" are output to the data output means (display device) 5, which is a computer display device.

In the above example, an example of determination using a prediction error (model error) is shown. However, the same concept can be applied to a case where a determination is made based on a correction coefficient or a correction amount. Further, instead of the determination based on the average error and the threshold value, it is possible to determine that a large error continues for a certain period.

According to the present invention described above, (a) when the error is small, that is, when the prediction model is appropriate, the prediction is performed while correcting the prediction value with the nearest error, and (b) when the error increases, If the characteristics of the data learned by the prediction model and the characteristics of the data at the time of the prediction have changed, the prediction model is re-learned to update the model. (C) If the error is further large, that is, if it is considered necessary to review the entire prediction model, a three-step model update based on the prediction error of changing the model configuration and reviewing the entire prediction model This is a method having means, and can perform such a series of processes automatically and adaptively.

ADVANTAGE OF THE INVENTION According to this invention, the adaptive prediction model construction method and the adaptive prediction model which enable high-precision prediction even when the time series data of a prediction target show an unexpected tendency, without imposing an unnecessary load on a system. A predictive model construction system can be provided.

For example, in the case of estimating power demand in a certain area, for example, when a new load (such as a factory) is newly established in the range of the estimation target, the conventional method cannot estimate the newly established load. Further, in the case of sales demand prediction as an example, the tendency of sales demand may change due to the advancement of other nearby stores, but the conventional method cannot cope with such a case.

電力 Electricity demand is different on weekdays, Saturdays and holidays, and power demand is different in different seasons. The same applies to various demands such as sales demands, water demands, and heat demands, since the trends in human social life are basically reflected. In order to predict such demand data, a single prediction model often cannot provide sufficient prediction accuracy. This is because it is too complicated to express the characteristics of the data with a single model. Therefore, it is necessary to construct a plurality of models and to make predictions for each prediction model for a period or section specialized for the model.

However, there has been no way to properly set this category, so a large amount of trial and error work was required. However, by using the present invention, it is possible to automatically perform appropriate prediction even when it is not known how to construct a model. Also, system developers, method developers, and users may have empirical knowledge on how to divide data.

For example, as mentioned earlier, power demand varies greatly on weekdays, Saturdays, and holidays, and power demand varies with seasons and temperatures. In the present invention, it is possible to use such an empirical data division as an initial state. This makes it possible to make the system easy for system developers, method developers, and users to understand. That is, the present invention can be applied to the case where the characteristics of the data to be predicted are unknown and the case where there is empirical knowledge about the data to be predicted.

Furthermore, in general, a certain amount of data is required to build a prediction model with high accuracy. Therefore, even if the operator recognizes the trend change of the prediction target time-series data, the prediction accuracy is not improved until data necessary for constructing a new prediction model is accumulated. According to the present invention, the prediction accuracy can be maintained by the predicted value correction unit until data necessary for updating the prediction model is accumulated.

It is a flowchart explaining the process of adaptive adaptive prediction model construction. It is a flowchart of a prediction model update step. It is a flowchart of a prediction model structure update step. It is a flowchart explaining a process of a prediction model construction step and a prediction model configuration update step. It is explanatory drawing of the degree of isolation using the center of a cluster. It is explanatory drawing of the degree of separation using the range of a cluster. FIG. 9 is an explanatory diagram illustrating generation of a cluster. FIG. 9 is an explanatory diagram illustrating correction of a cluster. It is a conceptual diagram explaining a multilayer neural network. 1 is a system configuration diagram of an adaptive prediction model construction system.

Explanation of reference numerals

1: Data input screen 2: Data input means 3: Data storage means 4: Database 5: Data output means (display device)
6: prediction model construction means 7: prediction model update means 8: prediction model configuration update determination means 9: prediction model configuration update means 10: alarm output means 11: prediction model 12: prediction execution means 13: prediction value correction execution means 14: Predicted value correcting means 100: network

Claims (14)

In an adaptive prediction model construction method for adaptively constructing a prediction model for predicting future time series data based on past time series data,
A prediction model construction step of constructing a prediction model using past time series data,
Performing a prediction using the prediction model, a prediction execution step of obtaining a predicted value,
A prediction value correction step of calculating a correction coefficient or a correction amount for correcting the prediction value based on a past prediction error or a model error,
Prediction error, model error, based on the correction coefficient or the correction amount, whether the model needs to be updated, or whether the model configuration needs to be updated, a prediction model update determination step of determining
When it is determined that model update and model configuration update are unnecessary, a predicted value correction execution step of correcting the predicted value,
A prediction model update step of updating the prediction model using the newly accumulated time-series data when it is determined that the model update is necessary;
A prediction model configuration update step of updating the prediction model configuration using the newly accumulated time-series data when it is determined that the model configuration needs to be updated;
A method for constructing an adaptive prediction model, comprising: The adaptive prediction model building method according to claim 1,
When the prediction error, the model error, the correction coefficient or the correction amount exceeds the allowable value for model update determination, it is determined that the prediction model needs to be updated, and the tendency / characteristics of the time series data to be predicted change. Alarm output step to output an alarm to notify that
A method for constructing an adaptive prediction model, comprising: The adaptive prediction model building method according to claim 1,
When the prediction error, the model error, the correction coefficient, or the correction amount exceeds the allowable value for model configuration update determination, it is determined that the prediction model configuration needs to be updated, and the trend / characteristics of the time series data to be predicted changes. Alarm output step to output an alarm to notify that
A method for constructing an adaptive prediction model, comprising: In the adaptive prediction model building method according to any one of claims 1 to 3,
The prediction model construction step and the prediction model configuration update step,
Analyzing time-series data of multiple factors, finding an input factor having an input-output relationship with a certain output factor from among the multiple factors, and modeling the input-output relationship,
A preliminary clustering step of performing clustering using all of the time series data to be used and dividing the cluster into preliminary clusters;
A clustering factor selection step of calculating the importance of the clustering factor for each preliminary cluster and selecting a clustering factor;
A clustering step of performing clustering using the time-series data related to the clustering factor selected by the clustering factor selection step to divide the cluster into clusters;
A factor determining step of determining an input factor and an output factor of the model for each cluster using the factors included in the cluster,
A modeling step of modeling the input / output relationship for each cluster using the time series data corresponding to the input factor and the output factor included in the cluster,
A model accuracy evaluation step of evaluating the accuracy of an integrated model integrating a plurality of models constructed for each cluster,
A cluster correction step of correcting the cluster so that the evaluation of the accuracy of the integrated model satisfies a predetermined condition;
With
An adaptive prediction model construction method characterized by sequentially performing a cluster correction step, a factor determination step, a modeling step, and a model accuracy evaluation step to build a model until the evaluation of the accuracy of the integrated model satisfies a predetermined condition. The adaptive prediction model building method according to claim 4,
The clustering factor selection step calculates a value representing the center of each factor for each of the preliminary clusters generated by the preliminary clustering, calculates the importance of the clustering factor based on the degree of dispersion of the values, and calculates the clustering factor. A method for constructing an adaptive prediction model characterized by selecting. The adaptive prediction model building method according to claim 4,
The clustering factor selection step calculates the importance of the clustering factor according to the degree of overlap of the range from the minimum value to the maximum value of each factor for each preliminary cluster generated by preliminary clustering, and selects the clustering factor. A method for constructing an adaptive prediction model characterized by the following. The adaptive prediction model building method according to claim 4,
The preliminary clustering step, the clustering factor selection step, and, instead of the main clustering step,
Adaptive prediction model construction characterized by subdividing in advance the time series data that can be used for the construction of a prediction model to create a plurality of time series data groups, and using this as an initial division to determine the model configuration Method. In an adaptive prediction model building system such as a computer that adaptively builds a prediction model for predicting future time series data based on past time series data,
Prediction model construction means for constructing a prediction model using past time-series data,
A prediction execution unit that executes prediction using the prediction model and obtains a prediction value;
Prediction value correction means for calculating a correction coefficient or a correction amount for correcting the prediction value based on a past prediction error or a model error,
Prediction error, model error, based on the correction coefficient or the correction amount, whether the model needs to be updated, or whether the model configuration needs to be updated, and a prediction model update determining means for determining
When it is determined that model update and model configuration update are unnecessary, predicted value correction executing means for correcting the predicted value,
Predictive model updating means for updating the predictive model using newly accumulated time-series data when it is determined that model update is necessary;
Predictive model configuration updating means for updating the predictive model configuration using newly accumulated time-series data when it is determined that the model configuration needs to be updated;
An adaptive prediction model construction system, comprising: The adaptive prediction model construction system according to claim 8,
When the prediction error, the model error, the correction coefficient or the correction amount exceeds the allowable value for model update determination, it is determined that the prediction model needs to be updated, and the tendency / characteristics of the time series data to be predicted change. Alarm output means for outputting an alarm for notifying that
An adaptive prediction model construction system, comprising: The adaptive prediction model construction system according to claim 8,
When the prediction error, the model error, the correction coefficient, or the correction amount exceeds the allowable value for model configuration update determination, it is determined that the prediction model configuration needs to be updated, and the trend / characteristics of the time series data to be predicted changes. Alarm output means for outputting an alarm for notifying that
An adaptive prediction model construction system, comprising: In the adaptive prediction model construction system according to any one of claims 8 to 10,
The prediction model construction means and the prediction model configuration updating means,
Means for analyzing time series data of a plurality of factors, finding an input factor having an input / output relationship with a certain output factor from among the plurality of factors, and modeling the input / output relationship,
Preliminary clustering means for performing clustering using all of the time-series data to be used and dividing the cluster into preliminary clusters;
Clustering factor selecting means for calculating the importance of the clustering factor for each preliminary cluster and selecting the clustering factor;
A clustering unit that performs clustering using the time-series data related to the clustering factor selected by the clustering factor selection unit and divides the cluster into clusters;
A factor determining means for determining an input factor and an output factor of the model for each cluster using the factors included in the cluster,
Modeling means for modeling the input / output relationship for each cluster using time series data corresponding to the input factor and the output factor, which are included in the cluster,
A model accuracy evaluation means for evaluating the accuracy of an integrated model obtained by integrating a plurality of models constructed for each cluster,
Cluster correction means for correcting the cluster so that the evaluation of the accuracy of the integrated model satisfies a predetermined condition;
With
An adaptive prediction model construction system, wherein a model is constructed by sequentially operating a cluster correction unit, a factor determination unit, a modeling unit, and a model accuracy evaluation unit until the evaluation of the accuracy of the integrated model satisfies a predetermined condition. . The adaptive prediction model construction system according to claim 11,
The clustering factor selection means calculates a value representing the center of each factor for each of the preliminary clusters generated by the preliminary clustering, calculates the importance of the clustering factor based on the degree of variation of the value, and calculates the clustering factor. An adaptive prediction model construction system, which is a means for selecting. The adaptive prediction model construction system according to claim 11,
The clustering factor selecting means calculates the importance of the clustering factor according to the degree of overlap of the range from the minimum value to the maximum value of each factor for each preliminary cluster generated by preliminary clustering, and selects the clustering factor. An adaptive prediction model construction system, characterized in that: The adaptive prediction model construction system according to claim 11,
In place of the preliminary clustering unit, the clustering factor selection unit, and the main clustering unit,
Adaptive prediction model construction characterized by subdividing in advance the time series data that can be used for the construction of a prediction model to create a plurality of time series data groups, and using this as an initial division to determine the model configuration system.

Images