JP2004086897A - Method and system for constructing model - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To efficiently construct an accurate model by automating the procedure required for modeling; and to flexibly follow a change of the application object by automatically adjusting the model form. <P>SOLUTION: This system is constituted such that input factors and output factors of the model are determined using factors included in each divided data space made by dividing multi-dimensional data space of a plurality of factors, modeling is performed using multivariate data in the respective divided data space, and the divided data space is corrected until conformation of accuracy of the integrated model made by integrating a plurality of models per the divided data space to a predetermined condition. <P>COPYRIGHT: (C)2004,JPO

Description

The present invention analyzes a multivariate data composed of a plurality of factors, finds an input factor having an input / output relationship with a certain output factor from a plurality of factors, and models the input / output relationship and a model. Concerning the construction system.

2. Description of the Related Art Analyzes for finding the input-output relationship lurking in multivariate data, modeling the data, and exploring data trends have been conventionally performed.
For example, a certain plant, or measurement information about the information processing system, production information, operation information, multivariate data consisting of multiple factors represented by failure information is obtained, the computer analyzes these multivariate data, plant -Modeling of various models typified by a prediction model of an information processing system or an abnormality diagnosis model.

Various modeling of the input / output relationship of the multivariate data can be considered. For example, a conventional technique for constructing a prediction model for predicting the future state of a certain factor in the multivariate data will be described.
It is assumed that the prediction model has an output factor and an input factor, and that the input / output relationship is determined using multivariate data corresponding to the input factor and the output factor.

In the method of constructing a prediction model, the following three types of methods are generally and often used.
(1) Method using past cases (2) Method using multiple regression equation (3) Method using neural network

As a method using the past case of (1), for example, Patent Document 1 (Japanese Patent Application Laid-Open No. 6-161989, title of “prediction device”) is disclosed. In this case, a case database is created by pairing the variables to be predicted and the values of variables that are considered to affect the prediction target, and cases similar to the input prediction variables are stored in the database. It is a method of taking out and making a prediction value. This method is often used when similarity or periodicity is observed as a property of a prediction target.

As a method using the multiple regression equation of (2), for example, Patent Literature 2 (Japanese Patent Application Laid-Open No. 6-174285, entitled "Air Conditioning Heat Load Prediction Method") is disclosed. This is a method of predicting the heat load by a linear regression equation of the maximum temperature, the minimum temperature, and the average humidity on the next day. As described above, the linear sum of the input factors used for prediction uses the multiple regression equation represented as the output factor, and is the most general prediction method.

The method of using the neural network of (3) is a method by which the neural network can model the nonlinearity between the input factor and the output factor, and has been particularly frequently used recently. For example, Non-Patent Document 1 (IEEJ Transactions B, Title: "Development of NN Applied Power Demand Forecasting System") can be cited.

Here, modeling of a neural network, which is a specific example of modeling, will be described with reference to the drawings. FIG. 11 is a conceptual diagram illustrating a multilayer neural network. Generally, a neural network has a multilayer neural network structure composed of an input layer, an intermediate layer, and an output layer as shown in FIG. 11, 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 the coupling between the elements of the neural network. If the coupling coefficient is large, the coupling is weighted, that is, it is determined to be a required coupling. If the coupling coefficient is small, the coupling is deemed to have no weight, 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 (multivariate 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.
A prediction model based on a neural network is like this.

The construction of such a prediction model is generally performed according to the following procedure.
(1) Analysis of multivariate data of output factors to be predicted (2) Determination of model configuration (3) Modeling (4) Model adjustment

In (1), various analyzes of multivariate data (hereinafter referred to as prediction target data) relating to output factors to be predicted are performed. In general, first, analysis is performed on what kind of tendency the prediction target data indicates, or Then, an analysis is performed to determine what correlation is found between the data to be predicted and the multivariate data of other factors, and the type of the prediction model to be applied is determined.

If strong periodicity is found in the data to be predicted, the use of past cases is considered to be effective.
If a linear correlation is found between the data to be predicted and the multivariate data of one or more other factors, the use of a multiple regression equation is considered effective.
When a nonlinear correlation is found or when a clear relationship cannot be found, the use of a neural network model is considered effective.

In (2), the model configuration is determined. Here, the model configuration determines whether to use a configuration in which the entire prediction target is predicted using one prediction model or a configuration in which prediction is performed using a plurality of prediction models.
For example, when a certain output of the above-mentioned plant is to be predicted throughout the year, a configuration in which prediction is performed using one prediction model, a configuration in which a prediction model is constructed for each season and prediction is performed, or For example, a method of constructing a prediction model for holidays and making a prediction is conceivable. In Non-Patent Document 1 cited earlier (IEEE Transactions B, titled "Development of NN Applied Electric Power Demand Forecasting System"), there are six types for spring, spring and summer, summer, summer and autumn, autumn and winter. Different types of prediction models have been built.

In the modeling of (3), a prediction model is created according to the results of (1) and (2). If a plurality of prediction models are used at that time, it is possible to specialize input factors suitable for each prediction model.

In the adjustment of the model in (4), a change such as addition or deletion of an input factor of a certain prediction model is performed using the prediction accuracy or the like as an evaluation index, or a change in each of a plurality of prediction models accompanying the review of the model configuration is performed. Do.
The construction of the prediction model was performed by such a method.
Although prediction has been described as an example of the conventional technology, the same work as described above is basically required for modeling of a model representing an input / output relationship according to monitoring, maintenance, abnormality diagnosis, and other purposes. is there.

JP-A-6-161989 (paragraphs 0014 to 0035, FIGS. 1 to 5) JP-A-6-174285 (paragraphs 0012 to 0019, FIGS. 1 to 4) 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"

Although the above procedure is generally performed, it is not actually an established modeling method. In the case of performing modeling, at present, an expert having expertise in the relevant field performs a certain period of time while performing each of the above operations (1) to (4) by trial and error by hand. There was a problem that a lot of man-hours until the end were required.

In addition, since specialized knowledge is required, there has been a problem that modeling work that satisfies satisfactory performance cannot be performed by non-experts.
Furthermore, since there is no unified framework applicable to various purposes, there is a problem in that work is performed while the above-described general procedure is embodied for each individual problem.

Furthermore, in a model such as a multiple regression equation or a neural network as described above, parameters inside the model are automatically calculated by selecting and determining input factors and output factors in advance and providing data.
However, since models such as multiple regression formulas and neural networks do not have a function of selecting input factors and output factors, even if an output factor is selected and an input factor unrelated to this output factor is selected, the input A model using factors is built. In such a model, there is a possibility that an improper model having a large error included in the output may be constructed.

Furthermore, when constructing an integrated model in which a plurality of models are combined, depending on the model configuration (for example, the number of models, etc.) and the structure of each model itself (for example, the number of elements of a neural network, etc.) constituting the integrated model. The model will be determined. The structure of the model configuration / model itself (hereinafter, referred to as a model form) is determined at the design stage, and thereafter treated as invariant. As the adjustment of the model form adaptive to the change of the target, only the adjustment of the parameters (regression coefficients in the multiple regression equation, weight coefficients of the neurons in the neural network) in each model was performed by the learning function.

However, when there is a large change in the application target, the model form cannot be automatically changed by the conventional method, and as a result of adjusting the parameters inside the model, if satisfactory performance cannot be obtained, There has been a problem that it is necessary to take measures such as adjusting the model form again by an expert or redesigning the system.

In general, in the above-mentioned conventional technology, there is no method for automatically performing modeling, that is, designing and changing the model form. Therefore, it is necessary for an expert to perform the operation manually, and to rely on the experience and intuition of the expert. In addition, there has been a problem that it is inevitable to repeatedly construct and adjust the model form, and it takes time to construct an optimal model.

The present invention has been made in order to solve the above-described problems, and an object of the present invention is to provide a method for automating a procedure necessary for the modeling, or an efficient method for manually performing the procedure required for the modeling. It is an object of the present invention to provide a model construction method and a model construction system that enable efficient construction of a high-precision model by supporting work.
Furthermore, a model construction method and a model construction system for automatically adjusting a model form (model configuration / structure of the model itself) to efficiently construct an adaptive model that flexibly follows changes in an application target. Is to provide.

In order to solve the above problems, a model construction method according to the invention described in claim 1 is
In a model construction method of analyzing multivariate 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 divided data space generation step of dividing a multidimensional data space represented by some or all of the factors to generate a divided data space;
A factor determining step of determining an input factor and an output factor of the model for each divided data space using the factors included in the divided data space,
A modeling step of modeling the input / output relationship for each divided data space using multivariate data included in the divided data space and corresponding to the input factor and the output factor,
A model accuracy evaluation step of evaluating the accuracy of an integrated model obtained by integrating a plurality of models constructed for each divided data space;
A divided data space correction step of correcting the divided data space so that the evaluation of the accuracy of the integrated model satisfies a predetermined condition;
Do
Until the evaluation of the accuracy of the integrated model satisfies a predetermined condition, a divided data space correction step, a factor determination step, a modeling step, and a model accuracy evaluation step are sequentially performed to construct a model.

The model construction method according to the second aspect of the present invention
In the model construction method according to claim 1,
The divided data space correction step corrects the divided data space by performing fusion / division of the divided data space in accordance with the presence or absence of a correlation using a correlation coefficient that is an index indicating the correlation between input and output of the model.
It is characterized by the following.

The model construction method according to the third aspect of the present invention
In the model construction method according to claim 1,
The divided data space correction step includes a step in which an error between a model output obtained by inputting multivariate data with respect to an input factor of a model determined for each divided data space and the multivariate data related to the output factor is a predetermined value. Divide the data above the threshold into a new data space and correct the divided data space,
It is characterized by the following.

The model construction method according to the fourth aspect of the present invention
In the model construction method according to any one of claims 1 to 3,
The divided data space generation step employs a clustering method of generating a cluster as a divided data space,
A preliminary clustering step of performing clustering using all of the multivariate data to be used and dividing 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;
This clustering step of performing clustering using the multivariate data related to the clustering factor selected by the clustering factor selection step and dividing the cluster into clusters,
Do
The method is characterized in that a multidimensional data space represented by some or all of the plurality of factors is divided to generate a divided data space in cluster units.

The model construction method according to the invention described in claim 5 is as follows.
The 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 model construction method according to the invention of claim 6 provides:
The 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.

The model construction system according to the invention described in claim 7 is:
In a model construction system such as a computer that analyzes multivariate data of multiple factors, finds input factors having an input / output relationship with a certain output factor from the multiple factors, and models the input / output relationship,
A divided data space generating means for generating a divided data space by dividing a multidimensional data space represented by some or all of the plurality of factors,
Factor determining means for determining an input factor and an output factor of the model for each of the divided data spaces using the factors included in the divided data space,
Modeling means for modeling an input / output relationship for each divided data space using multivariate data included in the divided data space and corresponding to the input factor and the output factor,
A model accuracy evaluation means for evaluating the accuracy of an integrated model obtained by integrating a plurality of models constructed for each divided data space;
Division data space correction means for correcting the division data space 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, the divided data space correction means, the factor determination means, the modeling means, and the model accuracy evaluation means are sequentially operated to construct a model.

The model construction system according to the invention of claim 8 is:
The model construction system according to claim 7,
The divided data space correction means corrects the divided data space by performing fusion / division of the divided data space according to the presence or absence of a phase using a correlation coefficient which is an index indicating a correlation between input and output of the model.
It is characterized by the following.

The model construction system according to the ninth aspect of the present invention
The model construction system according to claim 7,
The divided data space correction means is configured such that an error between a model output obtained by inputting multivariate data for an input factor of a model determined for each divided data space and multivariate data related to an output factor is a predetermined error. Divide the data above the threshold into a new data space and correct the divided data space,
It is characterized by the following.

The model construction system according to the invention described in claim 10 is:
In the model construction system according to any one of claims 7 to 9,
The divided data space generating means adopts a clustering method of generating a cluster as a divided data space,
A preliminary clustering means for performing clustering using all of the multivariate 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 multivariate data related to the clustering factor selected by the clustering factor selection unit and divides the cluster into clusters;
With
The method is characterized in that a multidimensional data space represented by some or all of the plurality of factors is divided to generate a divided data space in cluster units.

Further, a model construction system according to the invention described in claim 11 is:
The model construction system according to claim 10,
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 characterized by selecting.

The model construction system according to the invention of claim 12 is:
The model construction system according to claim 10,
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. It is characterized by.

According to the present invention as described above, a method of automating the procedure required for the modeling, or assisting to perform an efficient operation when manually performing the procedure required for the modeling, thereby improving accuracy. A model construction method and a model construction system capable of efficiently constructing a good model can be provided.
Furthermore, a model construction method and a model construction system for automatically adjusting a model form (model configuration / structure of the model itself) to efficiently construct an adaptive model that flexibly follows changes in an application target. Can be provided.

Hereinafter, the best mode of the present invention will be described. An embodiment of a model construction method according to claim 1 and an embodiment of a model construction system according to claim 7 will be described with reference to the drawings. FIG. 1 is a flowchart illustrating the flow of the model construction method, and FIG. 2 is an explanatory diagram illustrating the division of the data space.
The model construction method of the first embodiment analyzes multivariate data of a plurality of factors, finds input factors and output factors having an input / output relationship from the plurality of factors, and models the input / output relationship.

(1) Divided data space generation step (initial division)
Step S1 in FIG. 1 is a division data space generation for dividing a multidimensional data space represented by some or all of a plurality of factors of the multivariate data into a plurality of data spaces to generate a divided data space. Step. By the divided data space generation step, the multidimensional data space is divided into a plurality of data spaces including input factors, output factors, and related multivariate data having an input / output relationship.

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 division data space includes data as shown in the table of FIG. In the table of FIG. 2B, “□ (i)”, “× (i)”, “△ (i)”, and “○ (i)” represent some values other than 0 (zero). Shall be.
In the divided data space 1 and the divided data space 2, the values of the factors 1 and 2 are other than 0, and the values of the factor 3 are 0.
The divided data space 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.
The divided data space 4 is a space in which the values of the factors 1, 2, and 3 are other than 0.
The divided data spaces 1, 2, and 3 show a case where a divided data space whose value is 0 is generated by any factor.
The divided data spaces 1 and 2 are common when the value of the factor 3 is 0, but show a case where another divided data space is generated.
In such a divided data space generating means, for example, cluster analysis (clustering) and decision tree analysis can be applied.

Cluster analysis can be roughly divided into a hierarchical method and a non-hierarchical method.
In the hierarchical method, individual data in the multivariate data is regarded as one cluster, and a cluster having a large similarity (a small dissimilarity) is merged to generate a cluster. Although this cluster corresponds to a divided data space, in this specification, a divided data space generated by cluster analysis is particularly called a cluster for distinction. The cluster and the divided data space are the same and only differ in how they are called.

In the non-hierarchical method, a method of setting an initial cluster, comparing each data with the centroid of all clusters, relocating to the closest cluster, recalculating the centroid, and repeating this until there is no change in the centroid. It is.
By these cluster analysis, a plurality of clusters (divided data space) in which factors of multivariate data having a large similarity are put together are generated. Since this cluster does not include a factor having no similarity, it is easy to determine an input factor and an output factor in modeling.

Decision tree analysis is a method of expressing (modeling) rules (conditions) for classifying (predicting) multivariate data for a certain purpose in a tree structure. ID3 and its advanced forms C4.5, CART, CHAID is a typical method.
By this decision tree analysis, a plurality of divided data spaces in which factors of the multivariate data matching the conditions are collected are generated.

(2) Factor determination step Step S2 is a step of determining an input factor and an output factor of the model for each divided data space using the factors included in the divided data space.
By evaluating the generation result of the divided data space and the information on the division process of the divided data space, the input factor used in the modeling of (3) described later is determined.

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

決定 Also, in decision tree analysis, factors appearing at the top of the decision tree can be determined to be important factors. Therefore, it is possible to select an input factor from factors higher in the decision tree. Also, as a result of the decision tree analysis, a method that does not use variables that do not appear as branch conditions as input factors is possible.

(3) Modeling Step Step S3 is a modeling step of modeling an input / output relationship for each divided data space using multivariate data corresponding to input factors and output factors.
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 divided data space. 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 using an information criterion such as AIC (Akaike Information Criterion), and the like. Is applicable.

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

(4) Model accuracy evaluation step Step S4 is a model accuracy evaluation step of evaluating the accuracy of an integrated model obtained by integrating a plurality of models constructed for each divided data space.
The accuracy evaluation of the integrated model is performed by the following equation.

If the output error EE of Equation 1 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.

The multivariate data used for the model accuracy evaluation (error evaluation) includes a multivariate data evaluation method using the multivariate data obtained at the time of modeling as it is, and a multivariate data evaluation method for modeling. There is a method that does not use some of the multivariate data and saves them for model accuracy evaluation.
In the former method, it is possible to evaluate how much the multivariate data itself given for modeling has been modeled.
In the latter method, the generalization ability of the model can be evaluated.

(5) Divided data space correction step Step S5 is a divided data space correction step of correcting the data space so that the evaluation of the accuracy of the integrated model satisfies a predetermined condition.
More specifically, as a method of correcting the divided data space, a method of fusing a plurality of divided data spaces, a method of further dividing the divided data space to generate divided space data, and a method of replacing data in the divided data space are considered. Can be

In the method of fusing a plurality of divided data spaces, a fusion target divided data space determining step of determining a divided data space to be fused and a divided data space fusion step of fusing the divided data spaces are performed.
In the method of further dividing the divided data space to generate the divided space data, a separation target divided data space determining step of determining a divided data space to be separated, and data using any multivariate data in the target divided data space And a divided data space separating step of determining whether the data is a space.
Further, in the method of exchanging multivariate data in the data space, a replacement target data determining step of determining multivariate data to be exchanged and a relocation step of exchanging and relocating the multivariate data are performed.

(6) Model Reconstruction Step In steps S2 to S5, the divided data space 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 reconstruction step for constructing a dynamic model.
Evaluation of the accuracy of the integrated model, that is, when the output error EE of Equation 1 does not fall below a predetermined value, the divided data space is changed and included in the divided data space until the EE becomes less than a predetermined value. It repeats the model construction by the factors and multivariate data.

According to the present embodiment described above, a data space composed of factors of multivariate data is divided to generate a divided data space, and a data space including an output factor and an input factor related to the output factor is extracted. Since a model based on the output factor and the input factor is constructed based on the output factor and the multivariate data related to the output factor, a highly accurate model can be constructed.

Subsequently, a model construction system according to claim 7 will be described. FIG. 12 is a system configuration diagram of the prediction model construction system.
As shown in FIG. 12, the prediction model construction system according to the first embodiment 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 6, a divided data space. The apparatus includes a generation unit 7, a factor determination unit 8, a modeling unit 9, a model accuracy evaluation unit 10, and a divided data space correction unit 11.
Further, a network 100 such as the Internet, an intranet, and a LAN is connected to the data input unit 2. The model construction system corresponds to a computer (computer) on which these means function.

The multivariate data includes a data input screen 1 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 taken into the system through the input unit 2 and stored in the database 4 by the data storage unit 3. The data in the database thus accumulated is used. It should be noted that the data input means 2 can be performed through other means, for example, an electronic medium on which data is recorded.

(1) Divided data space generation means (initial division) 7
Means for generating a divided data space by dividing a multidimensional data space represented by some or all of the factors of the multivariate data into a plurality of data spaces, corresponding to step 1 in FIG. It has the function to do. As described with reference to FIG. 2, the divided data space generation means 7 divides the multidimensional data space into a plurality of data spaces including input factors, output factors, and related multivariate data having an input / output relationship.

(2) Factor determination means 8
This is means for determining an input factor and an output factor of a model for each divided data space using the factors included in the divided data space, and has a function corresponding to step S2.
By evaluating the generation result of the divided data space and the information on the division process of the divided data space, the input factor used in the modeling of (3) described later is determined.

(3) Modeling means 9
This is a means for modeling an input / output relationship for each divided data space using multivariate data corresponding to input factors and output factors, and has a function corresponding to step S3.
For modeling, typical methods are selected from the use of past cases, a multiple regression equation, a neural network, and the like.

(4) Model accuracy evaluation means 10
This is a means for evaluating the accuracy of an integrated model obtained by integrating a plurality of models constructed for each divided data space, and has a function corresponding to step S4.

(5) Divided data space correction means 11
This is a means for correcting the data space so that the evaluation of the accuracy of the integrated model satisfies a predetermined condition, and has a function corresponding to step S5.
More specifically, as a method of correcting the divided data space, a method of fusing a plurality of divided data spaces, a method of further dividing the divided data space to generate divided space data, and a method of replacing data in the divided data space are considered. Can be

In the method of fusing a plurality of divided data spaces, a fusion target divided data space determining means for determining a divided data space to be fused and a divided data space fusing means for fusing the divided data spaces are made to function.
In the method of further dividing the divided data space to generate the divided space data, the dividing target divided data space determining means for determining the divided data space to be separated, and the data based on any multivariate data in the target divided data space And a divided data space separating unit that determines whether the data is a space.
Further, in the method of exchanging multivariate data in the data space, the replacement target data determining means for determining the multivariate data to be exchanged and the relocation means for exchanging and relocating the multivariate data are made to function.

(6) Model Reconstruction Unit The divided data space correction unit 7, the factor determination unit 8, the modeling unit 9, and the model accuracy evaluation unit 10 are repeatedly operated until the evaluation of the accuracy of the integrated model satisfies a predetermined condition. It is a means to build a model.
If the evaluation of the accuracy of the integrated model does not fall below a predetermined value, the divided data space is changed and the factors and multivariates included in the divided data space until the output error EE of Equation 1 becomes smaller than the predetermined value. Repeat model building with data.

According to the present embodiment described above, a data space composed of factors of multivariate data is divided to generate a divided data space, and a data space including an output factor and an input factor related to the output factor is extracted. Since a model based on the output factor and the input factor is constructed based on the output factor and the multivariate data related to the output factor, a highly accurate model can be constructed.

Subsequently, a second embodiment of the model construction method / system according to claims 2 and 8 of the present invention will be described. The second embodiment is a more specific embodiment of the divided data space correction step / divided data space correction means of the first embodiment.
First, the divided data space correction step of the model construction method according to claim 2 will be described. The divided data space correction step S5 shown in FIG. 1 is performed by performing fusion / division of the divided data space in accordance with the presence or absence of a phase relationship using a correlation coefficient which is an index indicating the correlation between input and output of the model. This is the step of correcting the space.

In the method of correcting a divided data space according to the present embodiment, for multivariate data included in the divided data space, for example, an index representing a correlation of an input / output relationship such as a correlation coefficient is calculated. 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 divided data spaces C1, C2, and C3 are 0.6, 0.7, and 0.5, respectively.
Next, a combination of the divided data spaces is generated, and a new divided data space obtained by the combination is set as a temporary divided data space. In Table 1, three types of combinations are generated as combinations of provisional divided data spaces.
Subsequently, a correlation coefficient is calculated for the multivariate data included in the temporary divided data space. In the example of Table 1, four types of divided data spaces and their respective correlation coefficients were obtained together with the initial data space.

Here, it is determined which of the four data divisions is to be selected based on the correlation coefficient. As a method of the determination, for example, there is a method of selecting a division including a divided data space in which the correlation coefficient is maximum. In the case of Table 1, the provisional divided data space 3 is selected because it includes a data space 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 divided data space 1 is selected because the average of the correlation coefficient is 0.7.
Furthermore, it is conceivable that emphasis is placed on the correlation coefficient of the divided data space including multivariate data relating to a specific factor.

Such a divided data space correction step is performed until all the divided data spaces are finally merged and coincides with the original data space, and the overall accuracy of the integrated model is the highest among the divided data spaces obtained in the process. A high divided spatial data and its model are selected (note that such correction is defined as bottom-up correction in this specification and used in FIG. 4).

(5) A divided data space correcting means of the model construction system according to claim 8 will be described. The divided data space correction means of the second embodiment performs fusion / division of the divided data space in accordance with the presence or absence of a phase relationship by using a correlation coefficient which is an index indicating a correlation between input and output of the model, thereby obtaining a divided data space. And has a function corresponding to the divided data space correction step of the second embodiment described above.

Next, a third embodiment of the model construction method / system according to claims 3 and 9 of the present invention will be described. The third embodiment is a more specific embodiment of the divided data space correction step / divided data space correction means of the first embodiment.
The divided data space correction step S5 shown in FIG. 1 of the model construction method according to claim 3 is a method of obtaining a model output and an output factor obtained by inputting multivariate data with respect to an input factor of a model determined for each divided data space. And correcting the divided data space by performing fusion and division of the divided data space until an error from the multivariate data according to the above satisfies a predetermined condition.

In the third embodiment, an error in each model calculated in the accuracy evaluation of the entire model is used. 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, data in which this error is equal to or greater than a certain threshold is separated as a new separate divided data space.

Hereinafter, a specific example of the correction of the divided data space will be described with reference to the drawings.
FIG. 3 is an explanatory diagram for explaining the correction of the divided data space.
Although the data space is actually a multidimensional data space, FIG. 3 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 the divided data space generation step (initial division), the factor determination step, and the modeling step, the entire data space is divided into the divided data space 1, the divided data space 2, and the divided data space 3. The divided data space 1 includes six pieces, the divided data space 2 includes ten pieces, and the divided data space 3 includes five pieces of multivariate data.

モ デ ル Models 1, 2, 3 representing the input / output relationship of multivariate data for each divided data space are composed of neural networks, multiple regression equations, and the like. The model accuracy evaluation means described above inputs individual data in the multivariate data to the model 1, model 2, and model 3 corresponding to each divided data space, and calculates a modeling error. Here, it is assumed that the multivariate data when the modeling error is large (when it is larger than a separately set threshold) are circles (d1 to d7) in FIG.

多 The existence of such multivariate data indicates that modeling cannot be performed properly with the model for each divided data space at the present time. Therefore, the divided data space is corrected. As a method of correction, (A) a method of separating a new divided data space from an original divided data space, and (B) a multivariate data separated from each of the original divided data spaces is combined into one new divided data. A method of separating as a space is conceivable.

In the method (A), d1 and d2 are included in the new divided data space 1 ', d3, d4, and d5 are included in the new divided data space 2', and d6 and d7 are included in the new divided data space 3 '. That is, a divided data space is generated. In this case, the original divided data space 1 from which d1 and d2 are removed, the original divided data space 2 from which d3, d4, and d5 are removed, the original divided data space 3 from which d6 and d7 are removed, Will remain.
In the method (B), a new divided data space 4 having d1 to d7 as one is generated. In this case, the original divided data space 1 from which d1 and d2 are removed, the original divided data space 2 from which d3, d4, and d5 are removed, the original divided data space 3 from which d6 and d7 are removed, Will remain.

Such a divided data space correction step S5 is performed so that the divided data space is finally divided up to the set number, and the divided space having the highest overall accuracy of the integrated model is obtained from the divided data spaces obtained in the process. Data and its model are selected (note that such correction is defined as top-down correction in this specification and used in FIG. 4).

(5) A divided data space correcting means of the model construction system according to claim 9 will be described. The divided data space correcting means according to the third aspect is configured such that an error between a model output obtained by inputting multivariate data with respect to an input factor of a model determined for each divided data space and multivariate data relating to an output factor is predetermined. This is a means for correcting the divided data space by performing fusion / division of the divided data space until the condition is satisfied, and has a function corresponding to the divided data space correction step of the third embodiment described above.

Subsequently, a fourth embodiment of the model construction method / system according to claims 4 to 6, 10 to 12 will be described. The third embodiment is a more specific embodiment of the divided data space generation step and the divided data space generation means of the first embodiment.
First, a model construction method according to claims 4 to 6 will be described.
In the present embodiment, in the divided data space generation step S1 (initial division) described in the first embodiment, the data space is divided by a clustering method to generate a divided data space. Hereinafter, the divided data space will be referred to as a cluster, which is a name used in the clustering method, and will be described.
In the divided data space generation step in the present embodiment, as shown in FIG. 4, an initial setting step, a preliminary clustering step, a clustering factor selection step, and a main clustering step are further performed.

(1) Initial setting step S10
More specifically, the initial setting step S10 is a step for setting whether or not to execute clustering variable selection, setting a clustering method, setting a modeling method, setting a data division correction method, and setting other parameters.
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.
For the setting of the modeling method, a model such as a neural network or a multiple regression equation is selected.
The setting of the data division correction method determines whether the above-described method is based on the correlation coefficient or the analysis is based on error accuracy.
In the parameter setting, other necessary parameters other than the above (for example, the number of divided data spaces in the case of the model construction method and system according to claims 3 and 9) and other specific numerical values are input.

(2) Preliminary clustering step S11
The preliminary clustering step S11 is a step of performing clustering using all the multivariate data to be used and dividing the data into preliminary clusters.
The preliminary clustering is performed to obtain a preliminary cluster for obtaining the center of the cluster and the cluster range used in the following (3) clustering factor selection processing performed in the clustering factor selection step S12.

(3) Factor selection step for clustering S12
The clustering factor selection step S12 is a step of calculating the importance of the clustering factor for each preliminary 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 S12 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 Clustering Factor Based on Degree of Separation Using Cluster Center The following describes the degree of separation using the center of cluster. 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, a degree of separation using the center value of the cluster is defined for each factor, and a factor with a high degree of separation is selected as a factor for clustering (or a factor with 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, for example, the following equation (2).

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 as a result of performing clustering using two factors of the factor _{ｉ i} and the factor χ _j , 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, it can be determined that the more factors chi _i is an important factor in the degree of separation greater clustering than factor chi _j.

(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 Clustering Factor The selection of the clustering factor is performed 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 S13
The clustering step S13 is a step of performing clustering using the multivariate data related to the clustering factor selected in the clustering factor selection step S12 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 (data spaces).

Hereinafter, steps S14, S15, S16, and S17 in FIG. 4 correspond to steps S2, S3, S4, and S5 in FIG. 1, respectively. A similar method is used to construct a highly accurate model. Become.
When the model constructed in this way is used for prediction, diagnosis, relearning, or the like, the cluster to which the input data belongs is determined (determined by a distance function from the center of the cluster), and the model of the cluster is determined. Each operation (prediction, diagnosis, re-learning, etc.) is performed for this.

Further, a divided data space generating means according to the model construction system of claim 10 will be described. The divided data space correction means of the fourth embodiment further functions as an initial setting means, a preliminary clustering means, a clustering factor selecting means, and a main clustering means, and the initial setting means realizes a function corresponding to the initial setting step S10. The preliminary clustering means implements a function corresponding to the preliminary clustering step S11, the clustering factor selecting means implements a function corresponding to the clustering factor selecting step S12, and the present clustering means implements a function corresponding to the main clustering step S13. To achieve.
Then, the factor determining means 8 (corresponding to step S2), the modeling means 9 (corresponding to step S3), the model accuracy evaluating means 10 (corresponding to step S4), and the divided data space correcting means 11 (corresponding to step S5) shown in FIG. ) Function sequentially to construct a highly accurate prediction model 6.
Still further, the clustering factor selecting means of the model construction system according to claims 11 and 12 includes: (a) the degree of separation using the cluster center described above, as in the clustering factor selecting step S12 described above; (B) Selection is performed using the degree of separation using the cluster range.

When the prediction model 6 constructed in this way is used for prediction, diagnosis, relearning, and the like, the cluster to which the input data belongs is determined (determined by a distance function from the cluster center), and the cluster of the cluster is determined. Perform each operation (prediction, diagnosis, relearning, etc.) on the model.

Next, an example in which the fourth embodiment of the present invention is further embodied will be described using an application example to power demand prediction.
The power demand prediction is for predicting power consumption in a power system controlled by a power company, a specific area, a customer, and the like. The power demand to be predicted includes the daily amount of the next day or the day, the maximum power, the hourly power, and the like (in the present embodiment, the maximum power of the day is predicted). As a prediction model, a multiple regression equation, a neural network method, or the like is used. The power demand of the current day is predicted using the maximum temperature, the minimum temperature of the day, information indicating whether it is a weekday, a Saturday, or a holiday, and the power results of the previous day as input data.

Generally, the characteristics of power demand vary from season to season. That is, in summer, the power demand increases as the temperature rises (cooling demand), while in winter, the power demand increases as the temperature drops (heating demand). Also, the spring and autumn seasons are referred to as the interim periods, in which cooling and heating demands are small, and even if the temperature changes, the power demand does not change much. Electricity demand also differs greatly depending on weekdays, Saturdays and holidays. In general, power demand on weekdays is large, tending to be slightly less on Saturdays, and less on holidays.

As a prediction model for predicting power demand having such characteristics, when a single prediction model is used throughout the year, power demand having completely different characteristics is predicted by one model, and accuracy is reduced. Good predictions are difficult. Therefore, generally, a prediction model is constructed for each season, and a plurality of prediction model configurations are often used. However, there is a problem of how to determine the definition of the season, that is, how to determine the prediction model configuration, and it is difficult to obtain an optimal prediction model configuration.

Hereinafter, an embodiment of the present invention using power demand prediction as an example will be described with reference to FIG.
(1) Initial setting step S10
In the initial setting means, for example, the following contents are set.
Preliminary clustering method: longest distance method This clustering means: ward method Number of clusters to be divided: 4
Separation threshold used for selecting clustering factors: 0.1, 0.05
Divided data space correction step: bottom-up correction

(2) Preliminary clustering step S11
As preliminary clustering, the longest distance method, which is one of the hierarchical clustering methods, was applied, and divided into four clusters. In the hierarchical clustering method, cluster fusion is continued until one cluster is finally obtained. Therefore, the clustering is terminated when the number of clusters becomes four. Table 2 shows the factors used for clustering. Factors were implemented in two cases, Case 1 using five factors of maximum power, maximum temperature, minimum temperature, Saturday flag, and holiday flag, and Case 2 using three factors of maximum power, maximum temperature, and minimum temperature. .
* The Saturday flag and the holiday flag are binary data of 1 or 0, and are expected to have a very strong influence on the clustering result. For this reason, two cases were carried out, one using the Saturday flag and the other using the holiday flag.

予 備 The preliminary clustering results in each case are shown in FIGS. 7 and 8 are correlation diagrams of the maximum temperature and the maximum power to show the result of the preliminary clustering.

(3) Clustering factor selection step S12 (separation degree of claim 5)
A specific example in which the degree of separation using the center of a cluster is calculated as a selection of a clustering factor will be described with reference to the drawings. 9 and 10 are explanatory diagrams for explaining the centers of the clusters for each factor. FIG. 9A shows a bar graph, and FIG. 9B shows a table showing numerical values.
FIG. 9A shows, for each cluster in Case 1, the center value of each factor included in each cluster in a bar graph. The graph of the cluster 1 (C1) and the cluster 2 (C2) of the Saturday flag is not displayed because the center value is 0.

FIG. 9B shows the center value and the degree of separation (variance of the center of the cluster) of each factor included in each cluster for each cluster in Case 1. Here, the degree of separation (variance at the center of the cluster) is calculated using the above-described equation (2).

From FIG. 7 showing the clustering result, it can be seen that the maximum temperatures of all clusters are distributed from a range close to 0 (minimum value of maximum temperature) to a range close to 1 (maximum value of maximum temperature). That is, it indicates that the maximum temperature is not important for the clustering result.

On the other hand, looking at the degree of separation of the present invention shown in FIG. 9B, the maximum temperature and the degree of separation are very small values of 0.001 (the degree of separation of the minimum temperature is also very small as 0.002). Small), and can be determined to be less important for clustering than other factors. Looking at the Saturday flag, the center of cluster 1 and cluster 2 is 0, and the center of cluster 3 and cluster 4 is 1.

Since the value of the Saturday flag takes two values of 0 and 1, cluster 1 and cluster 2 represent days when the Saturday flag is 0 (that is, non-Saturday, weekday or holiday), and cluster 3 and cluster 4 represent Saturday. It can be seen that it represents. That is, it can be said that the Saturday flag (similar to the holiday flag) is an important factor for this clustering.

分離 Looking at the degree of separation shown in FIG. 9B, it is 0.333, which is the largest value among the factors, and it can be said that the degree of separation is effective as a method for finding a factor important for clustering.

Looking at Case 2, it can be determined from FIG. 8 that cluster rigging is performed according to the value of the maximum temperature. On the other hand, looking at the degree of separation in FIG. 10B, the value is 0.085, which is a large value as compared with other factors, and is consistent with the result determined from FIG.

Specific examples of the selection of the clustering factors will be described below.
In the case of Case 1, the separation degree of 0.1 was set as a threshold, and if it was less than 0.1, it was excluded from the clustering factors. That is, the three factors of the maximum temperature, the minimum temperature, and the maximum power are excluded, and only the Saturday flag and the holiday flag remain as clustering factors.

Similarly, in case 2, as a result of selecting a clustering factor with the threshold of the degree of separation set to 0.05, two factors, the highest temperature and the lowest temperature, remained as clustering factors. In the present embodiment, two types of preliminary clustering, Case 1 and Case 2, were performed, so that comprehensive judgment was made from the clustering obtained in each case, and the Saturday flag, holiday flag, maximum temperature, and minimum temperature were determined by four factors. Is a clustering factor.

(4) Main clustering step S13
Using the clustering factors (four factors of the maximum temperature, the minimum temperature, the holiday flag, and the Saturday flag) determined in the clustering factor selection step S12, this clustering was performed by the ward method, which is one of the hierarchical clustering methods.

FIGS. 14 to 18 show clustering results. The horizontal axis in the figure is the maximum temperature, and the vertical axis is the maximum power. FIG. 14 is a diagram showing the entire data before clustering is performed. 15 is a diagram showing cluster 1 (C1), FIG. 16 is a diagram showing cluster 2 (C2), FIG. 17 is a diagram showing cluster 3 (C3), and FIG. 18 is a diagram showing cluster 4 (C4).

(5) Factor determination step S14
Regarding the determination of the factors used in the prediction model, in this embodiment, the clustering factors (the four factors of the maximum temperature, the minimum temperature, the holiday flag, and the Saturday flag) used in this clustering are used as they are.

(6) Modeling step S15
As a modeling means, a three-layer neural network was used. That is, for each cluster, a neural network was constructed in which the maximum temperature, the minimum temperature, the holiday flag, and the Saturday flag were input and the maximum power was output.

(7) Divided data space correction step S17
With the four divisions obtained by the clustering means as an initial state, the division data space is corrected by bottom-up correction. FIG. 13 shows how the data space is corrected.
First, a combination that is divided into three is generated by fusing two clusters of the four divided clusters. In the present embodiment, five combinations of three divisions shown in FIG. 13 are generated. Here, in FIG. 13, the correlation coefficient (determined as a quadratic determination coefficient of the maximum temperature and the maximum power in consideration of the data distribution) is shown in the right column of the figure. The combination with the largest number was selected as the best case in the three divisions. That is, the division state of {C1}, {C2}, and {C3 C4} with a correlation coefficient of 0.610 is the best case in three divisions.

(4) Next, similarly, a combination for converting the three divisions into the two divisions is generated, and the correlation coefficient is evaluated. As a result, as shown in FIG. 13, three combinations of two divisions were generated. Among them, {C1} and {C2} C3 {C4} were obtained with the correlation coefficient of 0.635 as the combination having the largest correlation coefficient. This combination is the combination having the largest correlation coefficient among all the combinations, and is the final data division state obtained by bottom-up correction.

FIGS. 19 and 20 show correlation diagrams in the division {C1 {C2} C3 {C4} having the highest correlation. Finally, FIG. 21 shows learning errors as a result of learning the neural network model with {C1} and {C2} C3 {C4}. For comparison, a case where data division is not performed as a conventional method and a case of the present invention are both described. The horizontal axis of the figure represents days, and data for one year in total is shown. The vertical axis represents the error [%] between the actual value and the predicted value. According to this, it can be seen that the effect of the present invention is remarkably exhibited when the horizontal axis is in the range of 1 to 61, the horizontal axis is in the vicinity of 121, and the horizontal axis is in the range of 331 to 361. In the present embodiment, the prediction error of the present invention is about 8% higher than that of the conventional method between the prediction error of the conventional method and the prediction error of the present invention.

The prediction model construction method and the prediction model construction system have been described above.
As a first effect of the present invention, a data space is divided to generate a divided data space, input factors to the model are determined, and modeling is performed using multivariate data in the divided data space, and the entire model is Until the accuracy evaluation is satisfied, it is possible to automate the modeling operation by repeating the correction of the divided data space and the reconstruction of the model in the divided data space.

In addition, it is possible to provide assistance to a person who performs modeling work by using an interactive user interface for each phase without completely automating the processing flow shown here on a computer. . With this effect, the time required for the modeling operation can be reduced.

As a second effect of the present invention, conventionally, it was necessary to perform modeling work by a skilled person having a certain degree of expertise in the field, but by the present invention, it is possible to perform modeling even by an unskilled person in the field. Can be.

As a third effect of the present invention, by performing the above-described model construction periodically or at an arbitrary timing, the model is reconstructed adaptively when there is a change in the trend of the modeling target or a change in the trend. It is possible to go.

The method of adaptively adjusting a model has been conventionally realized by, for example, a neural network application system having a learning function, a multiple regression equation for executing multiple regression analysis at an arbitrary timing, or a Kalman filter application system.
However, in such a method, only the parameters used in the model being used are adjusted adaptively, and the model configuration itself is not adaptively adjusted.

As a specific example, in a system that uses a neural network to forecast demand for each season of spring, summer, autumn and winter, the designer performs the seasonal division of spring, summer, autumn and winter when designing the system, and the neural network is used for each season. Build the model. By implementing the learning function of the neural network, it is possible to adjust the neural network weighting coefficients of the prediction model for spring, the prediction model for summer, the prediction model for autumn, and the prediction model for winter. Adjustments such as shortening the end of spring and early start of summer are not possible.
In the present invention, it is obvious that the flexibility of the environment to be applied is high because it is possible to automatically determine the model section and adaptively adjust not only the model parameters but also the model section itself. is there.

As a fourth effect of the present invention, when generating a divided data space by clustering, by automatically selecting a factor to be used for clustering, or by giving an index which is a judgment for selecting a factor. is there. In a general clustering method, a cluster is generated using a similarity or a dissimilarity based on the Euclidean distance between data. Therefore, the result of clustering greatly differs depending on the factors used for clustering.

Also, if a factor that is not useful for clustering is used, the similarity including this factor will be calculated, which will be different from the original clustering result. However, by using the present invention, it is possible to eliminate factors that have an adverse effect on clustering in advance, and it is possible to perform appropriate data division, thereby improving the accuracy of a model finally constructed. Can be expected.

It is a flowchart explaining the flow of a model construction method. FIG. 3 is an explanatory diagram illustrating division of a data space. FIG. 9 is an explanatory diagram illustrating correction of a divided data space. It is a flowchart explaining the flow of a model construction method. 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. It is a correlation diagram of the maximum temperature and the maximum electric power for showing the clustering result. It is a correlation diagram of the maximum temperature and the maximum electric power for showing the clustering result. It is explanatory drawing explaining the center of the cluster for every factor, FIG.9 (a) is a bar graph, FIG.9 (b) is a table | surface which shows a numerical value. It is explanatory drawing explaining the center of the cluster for every factor, FIG.10 (a) is a bar graph, FIG.10 (b) is a table | surface which shows a numerical value. It is a conceptual diagram explaining a multilayer neural network. It is a system configuration diagram of a prediction model construction system. FIG. 9 is an explanatory diagram illustrating a relationship among the number of divisions, cluster division, and correlation coefficient. FIG. 9 is a characteristic diagram illustrating a clustering result in a relationship between maximum temperature and maximum power. FIG. 9 is a characteristic diagram illustrating a clustering result in a relationship between maximum temperature and maximum power. FIG. 9 is a characteristic diagram illustrating a clustering result in a relationship between maximum temperature and maximum power. FIG. 9 is a characteristic diagram illustrating a clustering result in a relationship between maximum temperature and maximum power. FIG. 9 is a characteristic diagram illustrating a clustering result in a relationship between maximum temperature and maximum power. It is a figure showing the correlation figure in division {C1 {C2} C3 {C4} with the highest correlation. It is a figure showing the correlation figure in division {C1 {C2} C3 {C4} with the highest correlation. It is a figure which shows the learning error of the result of having learned the neural network model by {C1} and {C2} C3 {C4}.

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 7: divided data space generation means 8: factor determination means 9: modeling means 10: model accuracy evaluation means 11: divided data space correction means 100: network

Claims (12)

In a model construction method of analyzing multivariate 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 divided data space generation step of dividing a multidimensional data space represented by some or all of the factors to generate a divided data space;
A factor determining step of determining an input factor and an output factor of the model for each divided data space using the factors included in the divided data space,
A modeling step of modeling the input / output relationship for each divided data space using multivariate data included in the divided data space and corresponding to the input factor and the output factor,
A model accuracy evaluation step of evaluating the accuracy of an integrated model obtained by integrating a plurality of models constructed for each divided data space;
A divided data space correction step of correcting the divided data space so that the evaluation of the accuracy of the integrated model satisfies a predetermined condition;
Do
A model construction method characterized by sequentially performing a divided data space correction step, a factor determination step, a modeling step, and a model precision evaluation step until an evaluation of the accuracy of an integrated model satisfies a predetermined condition, thereby constructing a model. In the model construction method according to claim 1,
The divided data space correction step corrects the divided data space by performing fusion / division of the divided data space in accordance with the presence or absence of a correlation using a correlation coefficient that is an index indicating the correlation between input and output of the model.
A model construction method characterized by the following. In the model construction method according to claim 1,
The divided data space correction step includes a step in which an error between a model output obtained by inputting multivariate data with respect to an input factor of a model determined for each divided data space and the multivariate data related to the output factor is a predetermined value. Divide the data above the threshold into a new data space and correct the divided data space,
A model construction method characterized by the following. In the model construction method according to any one of claims 1 to 3,
The divided data space generation step employs a clustering method of generating a cluster as a divided data space,
A preliminary clustering step of performing clustering using all of the multivariate data to be used and dividing 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 multivariate data related to the clustering factor selected by the clustering factor selection step to divide the cluster into clusters;
Do
A model construction method characterized by dividing a multidimensional data space represented by some or all of a plurality of factors to generate a divided data space in cluster units. The 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 variation in the value, and calculates the clustering factor. A model construction method characterized by selecting. The 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 model construction method characterized by the following. In a model construction system such as a computer that analyzes multivariate data of multiple factors, finds input factors having an input / output relationship with a certain output factor from the multiple factors, and models the input / output relationship
A divided data space generating means for generating a divided data space by dividing a multidimensional data space represented by some or all of the plurality of factors,
Factor determining means for determining an input factor and an output factor of the model for each of the divided data spaces using the factors included in the divided data space,
Modeling means for modeling an input / output relationship for each divided data space using multivariate data included in the divided data space and corresponding to the input factor and the output factor,
A model accuracy evaluation means for evaluating the accuracy of an integrated model obtained by integrating a plurality of models constructed for each divided data space;
Division data space correction means for correcting the division data space so that the evaluation of the accuracy of the integrated model satisfies a predetermined condition;
With
A model construction system wherein a model is constructed by sequentially operating divided data space correction means, factor determination means, modeling means, and model accuracy evaluation means until evaluation of the accuracy of an integrated model satisfies a predetermined condition. The model construction system according to claim 7,
The divided data space correction unit corrects the divided data space by performing fusion / division of the divided data space according to the presence or absence of a phase using a correlation coefficient that is an index indicating the correlation between input and output of the model.
A model construction system characterized in that: The model construction system according to claim 7,
The divided data space correction unit is configured such that an error between a model output obtained by inputting multivariate data with respect to an input factor of a model determined for each divided data space and multivariate data related to an output factor is a predetermined error. Divide the data above the threshold into a new data space and correct the divided data space,
A model construction system characterized in that: In the model construction system according to any one of claims 7 to 9,
The divided data space generating means adopts a clustering method of generating a cluster as a divided data space,
A preliminary clustering means for performing clustering using all of the multivariate 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 multivariate data related to the clustering factor selected by the clustering factor selection unit and divides the cluster into clusters;
With
A model construction system, wherein a multidimensional data space represented by some or all of a plurality of factors is divided to generate a divided data space in cluster units. The model construction system according to claim 10,
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 in the value, and calculates the clustering factor. A model construction system characterized by selection. The model construction system according to claim 10,
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. A model construction system characterized by:

Images