JPH08106448A - Weather forecasting device - Google Patents

Abstract

PURPOSE: To shorten the time of learning for a radar image which is newly measured and to find radar images at irregular forecasting time intervals with a small calculation quantity by systematically classifying and managing weather radar images and utilizing the weight of a neural network model which has learnt by using radar images by clusters. CONSTITUTION: Radar images from an input part 100 are classified by patterns by using the neural network for pattern classification (201). The weight of the neural network model which is generated in the past with a corresponding cluster and used by a learning and forecasting part 203 is obtained and set as an initial value for relearning. Further, optimum weight is generated for each pattern-classified cluster and registered in a data base part 202 so that it can be used. Further, an index for indicating forecasting time is inputted to the neural network model of the learning and forecasting part 203 together with the radar image.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention systematically classifies past radar images by applying a weather radar image to a neural network model for pattern recognition and learning to efficiently rainfall, snowfall, etc. The present invention relates to a meteorological feature amount, and a weather forecasting device that forecasts the number of products sold depending on the weather.

[0002]

2. Description of the Related Art First, an example of a neural network model to which the present invention can be applied will be given. Here, a hierarchical neural network model is used as a typical example, but it can be applied to other types of models such as a neural network model with regression coupling.

FIG. 2 is a diagram showing an example of a hierarchical neural network model. The hierarchical neural network model is a layered network model including one input layer, a plurality of intermediate layers, and one output layer, and each layer includes units, weights, and biases.

The unit is the output value (x
_i (i = 1, 2, ..., L, L: number of units in the previous layer) and weight (w _i , i: number of weight), and bias (b _i , i: unit number) It has a structure that receives the value obtained by adding as an input value, outputs a value (y) that is a non-linear transformation (f (•)) of the input value, and transmits this output value to the unit of the next layer (equation ( 1)).

However, here, the input / output conversion function of the unit of the input layer is linear, and the non-linear conversion function f (•) of the unit of the layer other than the input layer uses a sigmoid function which is a typical example (equation (1) However, it is also conceivable to use other conversion functions depending on the model.

Y = f (x) = 1 / {1-exp (−Σw _i x _i + b _i )} (1) [Σ is the sum from i = 1 to L] In the conventional weather forecasting device, A method has been proposed in which the radar image is predicted using the neural network model after learning by giving the measured radar image to the neural network model to learn the cloud movement (weather dynamics). For example, Japanese Patent Application No. 5-160530 “parallel calculation type weather radar image prediction device”, Japanese Patent Application No. 5-213830 “parallel calculation type rain radar image prediction device”, and Japanese Patent Application No. 5
-274065 "Nonlinear parallel calculation type rainfall radar image prediction device" gives the measured radar image to a neural network model having a product-sum calculation unit to train the meteorological dynamics, and the neural network after learning. The model is used to predict rainfall and snowfall.

However, no means has been used for systematically classifying the radar images based on the past radar images and utilizing them. Moreover, in order to make predictions in real time, it is necessary to further reduce the amount of calculation in learning of the neural network model, but no means for this is used.

[0008]

SUMMARY OF THE INVENTION The present invention is intended to systematically classify measured radar images based on pattern classification results of past radar images by automatically recognizing patterns of measured radar images. The purpose is to automatically generate an index that serves as a criterion. It also aims to shorten the learning time of the newly measured radar image by resetting the weights of the learned neural network model learned using the radar image belonging to the recognized cluster. .

Further, according to the present invention, based on the above-mentioned pattern recognition result of the radar image, learning is performed using the newly measured radar image and the radar images belonging to each classified cluster, and the neural network model model after learning is trained. The purpose is to facilitate the reuse of these data by registering the learning results such as weights in the database.

Furthermore, according to the present invention, by using a neural network model having a specific structure, it is possible to predict the weather at an arbitrary time with only one forward calculation and reduce the amount of calculation required for the prediction. To aim. It also aims to obtain radar images at unequal intervals of prediction time with a small amount of calculation.

[0011]

According to a first means of the meteorological forecasting apparatus of the first aspect, a neural network model is used for pattern recognition of the measured radar image.
This neural network model uses the radar image as an input and the similarity to each cluster that classifies the radar image as an output. The neural network model for pattern classification is created by learning with reference to the degree of similarity of each radar image determined by humans to each cluster. Various clusters of radar images can be defined according to the amount of clouds and the shape of clouds.

In the second means, the weight after learning of the neural network model belonging to the cluster having the radar image similar to the measured radar image is initialized based on the recognition result obtained by the first means. As a value, it is set in the neural network model for learning and retrained.

A third aspect of the weather forecasting apparatus according to the present invention.
In this method, a radar image belonging to a cluster having a radar image similar to the measured radar image is given to the neural network model for learning, and the weight after learning is registered as the optimum weight of the cluster. Or, the result of pattern recognition,
If the measured radar image does not belong to any cluster, a new cluster is added. At this time, the optimal weight of the neural network model for the added cluster is generated by the average value manipulation using the optimal weights of other clusters. Whether it does not belong to any cluster can be determined by comparing the maximum similarity with an existing cluster with a predetermined threshold.

According to a fourth aspect of the present invention, when the upper limit of the number of clusters (the number of clusters) when pattern-classifying a radar image is designated in advance, the learning function of the neural network model is used to determine the number of clusters. By changing it, the scale of the database containing the radar image and the weight value of the neural network model after learning is expanded or reduced. At this time, if a radar image having a pattern that is not registered in the database is input, a new cluster can be added to the specified upper limit.

When the scale of the database becomes too large, the number of clusters is reduced by integrating clusters having similar radar image patterns. In the fifth means of the weather forecasting device according to claim 4, in addition to the radar image as an input of the neural network model, an index representing the forecasting time is given, and as an output, the radar image after the forecasting time given at the input side is given. To get it.

In the sixth means, the radar image obtained by the fifth means is given again as an input value of the neural network model, and the radar image at a further previous time is predicted. By repeating this process, radar images at non-equidistant times are obtained by a small number of repeated calculations.

When a hierarchical neural network model is used as the neural network model used in the meteorological forecasting apparatus according to claim 4, the input / output characteristics of the unit in the middle layer are given by the following equation (Equation (2)).
It is described as. However, t is an index representing the prediction time of the radar image output by the neural network model.

Y = f (x, t) = 1 / {1-exp (−Σw _i x _i −w _{i + 1} t + b _i )} (2) [Σ is from i = 1 to L (L: previous) Total number of units per layer)

[0019]

According to the first and second means of the meteorological forecasting device of the first aspect, the pattern of the weather can be automatically identified by recognizing the pattern of the radar image. Further, by creating a database of past radar images based on the result of pattern recognition, the effort of systematically classifying and managing these radar images can be simplified.

The third and third aspects of the meteorological prediction device according to claims 2 and 3
In the means of No. 4, it is possible to automate the determination of the initial value of the weight given when re-learning the neural network model by using the above database. Also, the learning time required for re-learning is shortened because it is performed for each pattern-classified cluster.

In the fifth means of the meteorological prediction device according to the fourth aspect, it is possible to predict a radar image at an arbitrary time by giving an index representing a prediction time together with the radar image to the neural network model. . With the sixth means, it is possible to predict radar images at unequal time intervals with a small amount of calculation by repeatedly executing the fifth means.

[0022]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the weather forecasting device according to the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the weather forecasting apparatus according to claims 1 to 3, and 100 in the figure
Is an input unit, 200 is a data processing unit, and 300 is an output unit.

The input unit 100 comprises a weather radar 101 for measuring rainfall / snow areas and a file reading device 102 for reading information necessary for learning, pattern recognition and prediction of a neural network model.

The data processing unit 200 systematically classifies and manages the pattern recognition unit 201 for identifying the pattern of the measured radar image, the past radar image, and the optimal weight of the learned neural network model. Database part 2
02, a learning / prediction unit 203 for learning weather dynamics and predicting radar images.

A radar image measured by the weather radar 101 of the input unit 100 is input and transferred to the data processing unit 200. In addition, a value such as a learning rate necessary for learning is read from the file reading device 102 and transferred to the pattern recognition unit 201 and the learning / prediction unit 203.

The pattern recognition unit 201 inputs the radar image transferred from the input unit 100 into the pattern classification neural network model, and outputs an index for determining which cluster in the database the image belongs to.

The database unit 202 newly adds the radar image to a cluster having a pattern similar to the input radar image based on the determination result of the pattern recognition unit 201. Also, the optimal weights belonging to that cluster,
Also, the past radar image is transferred to the learning / prediction unit 203.

In the learning / prediction unit 203, the optimum weight transferred from the database unit 202 is set in the weather prediction neural network model as an initial value of the weight, and the radar image transferred from the input unit 100 and the database unit are set. 202
Relearn using the past radar image transferred from.
After retraining, fix the neural network model weights,
The radar image at the time after the radar image is measured is predicted, and the prediction result is transferred to the output unit 300. The output unit 300 displays the prediction result on a display or the like.

The neural network model used in the weather forecasting apparatus according to any one of claims 1 to 3 of the present invention uses, for example, a model as shown in FIG. 2, and handles radar images as input and output. On the other hand, in the neural network model used in the meteorological prediction device according to the fourth aspect, as shown in FIG. 3, information (t) regarding the predicted time is input to the input side. Based on this time information, the neural network model can arbitrarily change the time interval of the map formed between the input and the output (the time interval of the radar image between the input and the output).

FIG. 4 is a block diagram of an embodiment of the weather forecasting apparatus according to the present invention. In the figure, those having the same reference numerals as those in FIG. 1 are similar to those shown in FIG. 1, reference numeral 204 denotes an input switching unit for switching the radar image to be an input, and 205 denotes a prediction time designating unit for designating a prediction time.

In the case of this embodiment, the input switching section 2 is usually used.
Reference numeral 04 is set to select a radar image measured by the weather radar 101 as an input to the learning / prediction unit 203. The prediction time designation unit 205 provides the learning / prediction unit 203 with the information t regarding the predicted time together with the radar image. As a result, it is possible to predict not after a fixed time interval but after an arbitrary time according to the learned time. For example, when learning is performed by giving an index representing a prediction time of 5 minutes to a maximum of 30 minutes at the time of learning, the prediction result can be obtained by inputting once to the neural network model up to 30 minutes. When predicting a radar image at a time earlier than 30 minutes, the radar image after 30 minutes predicted by the learning / prediction unit 203 is given to the learning / prediction unit 203 via the input switching unit 204, and the required next By specifying the prediction time and repeating the prediction by the learning / prediction unit 203, the radar images at the times at unequal intervals are predicted with a small calculation amount.

In the weather forecasting device according to the first to third aspects, the mapping between the input and the output formed by the neural network model can be formed only at fixed time intervals. For example, when learning is performed so as to output a predicted radar image after 5 minutes with respect to the time of the input radar image, in order to predict the radar image after 30 minutes and 2 hours, respectively, 6 times, It is necessary to repeat the operation 24 times.

On the other hand, when the meteorological forecasting apparatus according to the fourth aspect is used, a model learned with a maximum forecast time of, for example, 30 minutes is used in order to forecast a radar image 30 minutes later and 2 hours later. In this case, the prediction can be performed only by repeating the operations once and four times, respectively.

In the above embodiment, for example, it is possible to predict the sales quantity of a product by learning the actual sales quantity of a product which is affected by the weather by a neural network model together with a radar image.

[0035]

With the first means of the meteorological prediction device according to the first aspect, it is possible to learn human judgment criteria when first creating a neural network model for pattern classification. It is possible to detect automatically based on a criterion close to human judgment criteria such as the similarity between the radar image and the past radar image.

Further, by the second means, based on the result of the pattern recognition by the first means, the weight after learning of the neural network model belonging to the cluster similar to the measured radar image is relearned. By resetting the network model, it is possible to automate the initial value determination of the weight and reduce the learning time required for re-learning.

In the meteorological prediction device according to the second aspect, the neural network model is constructed by using the third means, by using the past radar images systematically classified for the optimum weight for a specific cluster. It is possible to create without learning.

By using the fourth means according to the third aspect, it becomes possible to change the number of clusters in the pattern classification of the radar image by using the learning function of the neural network model. , It is possible to automatically expand or reduce the scale of the database that manages the weight values of the neural network model after learning.

In the meteorological prediction device according to the fourth aspect, it is possible to predict the radar image at any time by using the fifth means. Further, since the prediction can be made by only one forward calculation, it is possible to reduce the amount of calculation required for the prediction and the prediction processing time. Further, by using the sixth means, it is possible to repeatedly predict radar images at arbitrary times, and thus it is possible to obtain radar images at predicted times at unequal intervals with a small amount of calculation.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of a weather forecasting apparatus according to claims 1 to 3.

FIG. 2 is a diagram showing an example of a neural network model used in the weather forecasting device according to any one of claims 1 to 3;

FIG. 3 is a diagram showing an example of a neural network model used in the meteorological prediction device according to claim 4;

FIG. 4 is a block diagram of an embodiment of a weather forecasting device according to claim 4;

[Explanation of symbols]

 100 input unit 101 weather radar 102 file reading device 200 data processing unit 201 pattern recognition unit 202 database unit 203 learning / prediction unit 300 output unit

Claims (4)

[Claims] 1. A meteorological prediction device for providing a weather radar image to a neural network model to learn weather dynamics, and using the neural network model after learning to perform short-term prediction of weather such as rainfall and snowfall, (1) First means for applying the measured radar image to a neural network model for pattern classification, performing pattern recognition based on a classification standard of past radar images, and classifying the radar images by pattern (2) ) Based on the pattern recognition result of the measured radar image, the weight of the neural network model for weather prediction created using past radar images having a pattern similar to the pattern of the measured radar image is set as the initial value. A weather forecasting device comprising: a second means for setting the weather forecast neural network model and re-learning. 2. The weather forecasting device according to claim 1,
Radar images belonging to a specific cluster that are pattern-classified by the first means are given to a neural network model for weather prediction for learning, and the clusters are represented by the clusters in order to represent the weather dynamics contained in these radar images. It has a third means for creating optimum weights for the clusters, or for creating new clusters if they do not belong to any cluster and for generating optimum weights of the neural network model according to the clusters. A weather forecasting device. 3. The weather forecasting device according to claim 2,
Based on the result of the pattern recognition by the first means, the weight of the neural network model created by the third means and the radar image used for learning are registered and managed in a database, and a cluster is added. Alternatively, the weather forecasting device is provided with a fourth means for automatically enlarging / reducing the scale of the database by integration. 4. The weather forecasting apparatus according to claim 1, claim 2 or claim 3, wherein (1) a radar network image is provided to a neural network model for weather forecasting, and an index representing a forecasting time is given to an arbitrary A fifth means for predicting a radar image at a time and (6) a sixth means for predicting a radar image at an arbitrary arbitrary time by repeatedly giving the predicted radar image to a neural network model are provided. A meteorological prediction device having.