JP3005141B2 - Weather forecast system and neurocomputer control system - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a weather forecasting technique and a neuro-computer control technique, and more particularly to a technique effective for a weather forecasting service such as the Japan Meteorological Agency.

[0002]

2. Description of the Related Art Conventional weather prediction is performed by using a supercomputer based on weather information (atmospheric pressure, temperature, humidity, wind direction, wind speed, etc.) sent from each weather observation point. Calculates future atmospheric conditions (atmospheric pressure, calorific value), and the forecaster uses the past experience to determine "clear"
A method of judging a specific weather such as "rain" was generally used. Recently, a study has been conducted in which a neurocomputer learns the atmospheric pressure arrangement and the weather under the conditions, inputs the predicted atmospheric pressure arrangement at the predicted time to the neurocomputer, and outputs the weather corresponding to the atmospheric pressure arrangement at that time. . In addition, the rainfall probability causes the computer to output "statistical information that in such past atmospheric conditions, in the past, it would rain so much".

[0003] As related to the above-mentioned prior art, Asahi Shimbun '92 .4.21 "Anything Q & A: Weather Forecast (1)" Japanese Patent Application Laid-Open No. 4-220758 discloses "Time series data prediction and prediction recognition." There are techniques disclosed in literature such as "Method".

[0004]

The above prior art has the following problems.

[0005] That is, first, the past atmospheric pressure arrangement and the weather under the situation are made to be learned by the neurocomputer, and the predicted atmospheric pressure arrangement at the predicted time is input to the neurocomputer, and the weather at that time is output. In the case, for example, when it is at the center of a mobile high pressure,
However, there is no problem, but the probability of hitting is greatly reduced in the case of a complex air pressure arrangement, in the case of a drastic change, or at the time of a seasonal change.

[0006] Second, in numerical forecasting, in which dynamic calculations are performed using data such as atmospheric pressure and calorific value to estimate future atmospheric conditions, large devices such as supercomputers and calculations are executed. You need a huge program for.

Third, in order to obtain the probability of precipitation, when the computer outputs "statistics that in the past, in the case of such atmospheric conditions, it would rain so much," statistical information corresponding to every case is output. Need, and therefore a huge amount of data.

Fourth, in the large-scale weather forecasting system as described above, it is difficult to perform weather forecasting (such as subsequent weather, precipitation probability, etc.) without specialized knowledge of weather, not to mention the size of the apparatus. It is. Therefore, even if it is a local weather forecast, it is necessary to ask a specialized trader.

In the present invention, the atmospheric weather condition is strongly time-series, and generally the atmosphere moves slowly from west to east,
By taking advantage of the fact that even in exceptional typhoons, movement and atmospheric conditions are regular. The neural computer stores past weather information (atmospheric pressure, temperature, humidity, wind direction, wind speed, precipitation probability, clearing probability,
The purpose is to learn changes in the probability of snow, the probability of heavy rain, etc.), then input previous time-series weather information, and infer subsequent weather information.

It is a further object of the present invention to improve the accuracy of weather prediction by learning and estimating the state of changes in the atmospheric condition every year for each seasonal factor.

[0011] In addition to predicting the nationwide version of the weather information provided by the nationwide version of the weather information, local weather prediction of any area can be performed by adding the weather information of the area where the weather is to be predicted to the learning target of the neurocomputer. The purpose is to.

Further, in making a local prediction of an arbitrary region, the fact that the correlation between the "specific region" and "the state of the change in the atmospheric state of the specific region" is strong, the specific region is used. It is intended to perform subsequent weather forecasts simply by performing weather observations.

[0013] The second that neurocomputer guessed
Weather information (precipitation probability, sunny probability, snow probability, heavy rain probability, etc.) by threshold operation, "1" (positive factor),
Or, by converting the weather information input to the neurocomputer into a value of “0” (negative factor) and digitizing the case where the regularity is the highest to the case where the regularity is the lowest, the input weather information is converted. The purpose is to investigate "fractal dimension of temporal change" and estimate the probability of a positive factor of the second weather information from 100% to 0%.

Further, the present invention provides: The purpose is to accurately perform weather forecasts without extensive equipment and specialized knowledge.

[0015] A neurocomputer (an aggregate of neurochips or an arithmetic unit using a CPU such as RISC) and a neuroboard composed of its control device are incorporated in a computer system, and an individual system for controlling the neurocomputer is controlled by an application on an operating system. By providing a subroutine, for example, "weather forecast system" described in the present invention,
An object of the present invention is to allow a user to incorporate various processes for controlling a neurocomputer into an application according to the purpose.

The above and other objects and novel features of the present invention will become apparent from the description of the present specification and the accompanying drawings.

[0017]

SUMMARY OF THE INVENTION Among the inventions disclosed in the present application, the outline of a representative one will be briefly described.
It is as follows.

That is, according to the first aspect of the present invention, weather information indicating the atmospheric state of a plurality of points for performing weather observation is collected at a plurality of times, and one or more of the one or more points among the plurality of points are collected. A weather forecasting system for estimating weather information as described above, wherein a first control means for inputting weather information indicating atmospheric conditions at a plurality of points for performing weather observation collected at a plurality of first times to the weather forecasting system. And using a neural computer to estimate and output weather information at one or more points among the plurality of points at a second time after the first time using learning information based on the weather information. 2
And third control means for comparing the estimated value at the second time and the observed value at the second time to update the learning information.

According to a second aspect of the present invention, in the weather forecasting system of the first aspect, the weather forecasting system identifies a plurality of patterns for classifying the atmospheric condition.
Control means, and for each of the patterns, a fifth control means for generating a learning information by learning a state of a change in the atmospheric condition by a neurocomputer, wherein the second control means corresponds to the pattern This is to estimate weather information using learning information.

Further, the invention according to claim 3 comprises an arithmetic unit for executing a weather forecasting program and performing data processing, a memory storing the weather forecasting program, and an internal bus connecting the arithmetic unit and the memory. A neuro-computer comprising a neuro-board having an information processing device, a neuro-computer, and a neuro-computer controller connected to an internal bus of the information processing device for controlling the neuro-computer. The control device is
Receives weather information collected at a plurality of points at a plurality of times at which a weather observation is performed from an information processing apparatus executing the weather prediction program, and, based on the weather information, after the first time using learning information. Estimating the weather information at one or more points of the plurality of points at the second time, and comparing the learning information by comparing the estimated value at the second time with the observation value at the second time. It controls updating.

According to a fourth aspect of the present invention, in the weather forecasting system according to the first or second aspect, or in the neuro-computer control system according to the third aspect, observation and estimation are performed at a plurality of times and at a plurality of points. A first step of allocating necessary weather information to the input and output neurons of the neurocomputer, and a plurality of points at a plurality of times before an arbitrary time,
And, the second stage of inputting the content of weather information observed at one or more points at one or more times after the time to the neurocomputer, based on the content of weather information before any time, By feeding back the error between the content of the weather information after the arbitrary time estimated by the neurocomputer and the content of the weather information observed after the arbitrary time, the learning information (connection weight, The third step of repeating the operation of updating the threshold value) until the error is sufficiently small, and by inputting weather information of each point at a plurality of times before the current time into the neurocomputer using the learning information. A fourth function of outputting necessary weather information at each point after the current time, and a first function having a seventh control means for performing a control operation. The eighth control means for learning the state of the change in the atmospheric state of each point independently for each of at least one or more points where the necessary weather information is observed, and the neuro computer, The ninth control means for learning the state of the change in the atmospheric state at an arbitrary point selected from the respective points corresponding to the change, and the neuron for each point using the learning result of the eighth control means. Tenth control means for inputting weather information at a plurality of times before the current time of a point to a computer and outputting necessary weather information at at least one or more different times after the current time, and ninth control means Using the learning result of the means, the neurocomputer for each arbitrary point selected by the ninth control means,
A second function, comprising: an output result of the tenth control means, and an eleventh control means for outputting necessary weather information at at least one or more different times after the current time. Twelfth control means for learning independently the change of weather information at an arbitrary point where necessary weather information is observed among the information for each weather information; A thirteenth control means for learning a state of change of one piece of weather information selected from each piece of weather information at an arbitrary point corresponding to the change,
Using the learning result of the twelfth control means, for each piece of weather information, input one piece of weather information of at least one or more different times before the current time at an arbitrary point, and at least one or more pieces of weather information after the current time. The fourteenth control means for outputting one weather information at different times and the learning result of the thirteenth control means, for each arbitrary weather information selected by the thirteenth control means, a neuro computer, An output result of the fourteenth control means is inputted, and one weather information of at least one or more different times after the present time at an arbitrary point is output.
And at least one of the third functions of the control means.

According to a fifth aspect of the present invention, there is provided a weather forecasting system according to the first or second aspect, or the neurocomputer control system according to the third aspect, wherein the neurocomputer is provided with a weather forecast at an arbitrary point for performing weather observation. Using a learning result of the sixteenth control means and a learning result of the sixteenth control means, the information learns a state of change of one weather information among weather information such as a probability of precipitation, a probability of clearing, a probability of snow, a probability of heavy rain, and the like. Seventeenth control means for inputting one weather information at a plurality of times before the current time at a point to the neurocomputer and outputting one weather information at one or more times after the current time, and chaos of the input weather information A fourth function consisting of an eighteenth control means for examining the probability of hitting the weather information output by the seventeenth control means. The weather information output by the data is converted into "1" or "0" by threshold operation, and the weather information is divided into positive factors and negative factors. The fractal dimension of the temporal change of the input weather information is examined by using a mechanism that quantifies the case where the regularity is the lowest, and the probability that the positive or negative factor of the estimated weather information falls between 100% and 50%. Calculated by the value of
The probability of a positive or negative factor in weather information being 1
Fifth function consisting of nineteenth control means and sixteenth and seventeenth control means obtained by a value between 00% and 0%, the weather information output by the neurocomputer in the eighteenth control means is converted into weather information The firing rate for the affirmative factor is calculated, and the fractal dimension of the temporal change of the input weather information is examined using a mechanism that quantifies the case where the regularity of the data is the highest to the case where the regularity of the data is the lowest. The twentieth control means and the first control means determine the probability of the positive factor of the information as a value between 100% and 50%, and obtain the product of the firing rate and the probability of the positive factor of the estimated weather information as the weather information.
And a sixth function comprising a sixth control means and a seventeenth control means.

[0023]

The following is an example of the operation of the above-described weather forecasting system and neuro-computer control system of the present invention.

(1) Weather forecasting system I
A weather forecast system according to claim 2 or claim 4, and a weather forecast system according to claim 1.
First, second, and seventh control means described in (1) The present system allows the neurocomputer to learn the state of changes in the atmospheric state at a plurality of observation points, and obtain time-series weather information before the current time at each observation point. The information is input to infer weather information at one or more different times after the present time at each observation point. The "neurocomputer basic model in the weather forecasting system I" shown in FIG. 4 and the "neural computer basic model shown in FIG. An explanatory diagram showing an example of the relationship between neurocomputer input and output "will be used.

Time T for performing weather observation and weather forecast
(t ₀₎ weather information F _in the input of each observation point L (j) corresponding to _{(t 0, j, i 0} ) (F in (i 0 as an example): atmospheric pressure, temperature, humidity, wind direction , Wind speed) and weather information F _out (t ₁ , j, i ₁ ) for output (for example, F _out (i ₁ ): atmospheric pressure, temperature, humidity, wind direction, wind speed, rainfall probability, clearing probability) Each neuron N in the input layer of the computer
_in (t ₀ , j, i ₀ ) and the output layer neuron N _out
(t ₁ , j, i ₁ ), and the observed weather information F _in (t ₀ , j, i ₀ ) that is actually observed at each observation point and collected by an information processing device (hereinafter referred to as a host) So that the upper and lower limits of the observable range are real numbers from [0] to [1].
Transformed using the sigmoid function, a further 2 ⁿ multiplied by the value a meteorological measured information _{H in (t 0, j,} i 0),. First, the weather information of each observation point observed at different times (for example, 24, 12 and 6 hours before and at the present time) is converted into weather measurement information and input to the input layer of the neurocomputer, and the learning information (neuro Using the computer's hidden layer, the threshold value of each neuron in the output layer, and the coupling coefficient stored in a plurality of synapses, the weather condition of each observation point thereafter (for example, after 6, 12, 24 hours) Is estimated, and the weather estimation information H _out (t ₁ ,
j, i ₁₎ is outputted from the output layer as then the relative weather estimation information H _out, was observed actually relevant time weather information _{F in (t 1, j,} i 1) weather actual information H _in Is input to the output layer of the neurocomputer to obtain an error between the weather estimation information H _out (t ₁ , j, i ₁ ) and the weather measurement information H _in (t ₁ , j, i ₁ ). Is updated to minimize the learning information; By repeating the item numbers, the state of weather change (one event) is learned, and further, a plurality of "states of weather change" (n events) at different times are learned. For example, 2 sent from each observation point to the host
4,12,6 hours ago, and the current weather information F _in (t
_{0: k0~k1, j, i 0} ) weather actual measurement information H _in (t _0, j,
i ₀ ), input to the input layer of the neurocomputer, and infer the subsequent weather condition using the learning information, and output from the output layer as weather estimation information H _out (t ₁ , j, i ₁ ). And the weather information F _out (t ₁ : k1 to k2, j, i
_By converting to ₁ ), weather information at each observation point L (j) after 6, 12, and 24 hours is estimated, for example.

From the above,. By having the neurocomputer learn past weather information, and then input previous time-series weather information,
The subsequent weather information can be estimated.

[0027] In the present invention, since observation and prediction points of weather information can be arbitrarily set, a nationwide version of weather information is provided. By adding it to the learning target, local prediction of an arbitrary region is possible.

[0028] The target of neuro computer learning,
Probability of precipitation, probability of sunny, probability of snow, probability of heavy rain,
By adding at least one or more weather information, such information can also be predicted.

[0029] The operation of the weather forecasting system according to the present invention does not require specialized knowledge on weather.

(2) Judgment of Seasonal Factors (Third Control Means According to Claim 1, Fourth and Fifth Control Means According to Claim 2) The neurocomputer sends a change in atmospheric state and a change in seasonal factors. Learning the relationship, inputting time-series weather information, estimating the seasonal factors of the current atmospheric condition, and using the learning information corresponding to the seasonal factors (learning results of changes in atmospheric conditions) Make predictions.

The state of the change in the atmospheric condition is shown in spring, rainy season, summer,
Seasonal factors such as typhoons, autumn, long rains in autumn, and winter are targeted for learning, and preprocessing for learning and estimating is as follows. The input weather information F for each observation point L (j) corresponding to the time T (t ₀ ) at which weather observation and weather prediction are performed.
_in (t ₀ , j, i ₀ ) is used for each neuron N _in (t ₀ , j, i ₀ ) of the input layer of the neurocomputer, and for the spring, rainy season, summer,
Each seasonal factor such as typhoon, autumn, long rainfall in autumn, winter, etc. is assigned to the neuron N _out (seasonal factor) in the output layer, and observed at different times (for example, 24, 12, 6 hours before and at the present time). The weather information of the observation point is converted into actual measurement information and input to the input layer of the neurocomputer.
Using the seasonal factor discriminating learning information (in the middle layer of the neurocomputer, the threshold value of each neuron in the output layer, and the coupling coefficient stored in a plurality of synapses), the seasonal factor for the weather information is represented by [0 ] From the output layer in real numbers from [1] to [1]. Next, the most appropriate seasonal factor for the weather information is correct [1], and the non-applicable seasonal factor is [0]. By updating the learning information used by the neurocomputer so that the `` error between the seasonal factor determined by the neurocomputer and the seasonal factor input to the output layer as a correct answer '' is minimized by inputting to the layer, . By repeating the item numbers, the seasonal factor (one event) corresponding to the change in weather is learned, and a plurality of “seasonal factors corresponding to the change in weather” at another time (n events)
To learn about. As preprocessing for actually estimating weather information, for example, 24, 1 sent from each observation point to the host
Weather information F _in (t ₀ : k0) two to six hours ago and the current time
～K1, j, converts i) weather actual measurement information _{H in (t 0, j,} i) , the input to the input layer of the neural computer, by using the learning information for seasonal factors determine, seasonal factors for the weather _Is output (determined) from each neuron N _out (seasonal factor) of the output layer 22 with a real number from [0] to [1], and the season assigned to the neuron N _out (seasonal factor) that has output the largest value is The factor is determined as a seasonal factor with respect to the weather information, and based on the “learning information obtained by learning the state of weather change for each seasonal factor”, for example, 6,12 , 24 hours later, weather information F _out (t ₁ : k1 to k2, at each observation point L (j))
j, i).

As described above, the range of learning for the weather by the neurocomputer can be limited to “seasonal”, weather conditions having a strong correlation in a time series can be learned, and subsequent weather conditions can be determined and finely estimated. Prediction accuracy at the turn can be maintained, and the power can be demonstrated.

(3) Neurocomputer Control System (Neurocomputer Control System and Sixth Control Means According to Claim 3) For an application running on an OS (Operating System), general-purpose for controlling the neurocomputer on the OS Provides a typical subroutine.

This system runs on the OS of the information processing device,
For each process that controls the neurocomputer for the application, for example, "learning process" or "guessing process" which is a basic function, or "allocation of neurons to be learned" which is a mode setting of the neurocomputer, "learning information" And "determination of the number of neurons in the input / output layer and the intermediate layer" are provided as subroutines.

On the other hand, in the course of executing the program, the application sends the above command to the OS as necessary.
(A subroutine is started on the OS).

In response to the command, the OS transmits a request for an application to the neurocomputer control system, and the neurocomputer control system prepares control information for controlling the neurocomputer as required for the request, Requests the neurocomputer to execute the instructions. The neurocomputer control system waits for an end report for the instruction of the neurocomputer, reports the end to the OS, and stores information to be provided to the application or the OS in a predetermined location as necessary. Upon the completion report, the OS executes a completion report to the application or a request from the neurocomputer control system.

As described above, the information processing apparatus is connected to the "neurocomputer control device to which the neurocomputer is connected" by the internal bus or the standard interface.
By incorporating a subroutine provided on S into an application program, processing using a neurocomputer can be executed.

(4) Weather forecast system II (claim 1,
The weather forecasting system according to any one of claims 2 and 4, and the eighth, ninth, tenth, and eleventh control means according to claim 4. Learn how the state changes,
Further, by learning the state of the change of the atmospheric state of a specific point with respect to the change of the atmospheric state of each point, the weather information of each observation point at a different time before the current time is input, and the one after the current time of the specific point is input. This is for estimating weather information at two or more different times, and will be described with reference to a "neurocomputer basic model in weather forecast system II" shown in FIG.

[0039] At any observation point L (0), at least one or more different times T (t ₀ ), the necessary weather information of the weather information observed and estimated is transmitted to the input layer and the output layer of the neurocomputer. Assigned to each neuron,. In the point L (0), an arbitrary time T (t ₀₎ before, and at least one different time of the time T (t ₀₎ and later, the observed content of the weather information, the neuro-computer To learn the state of the change of the atmospheric state at the point L (0); Executing the item numbers at at least one or more points L (j) for performing weather observation; At each of the points L (j) where the item numbers are executed, necessary weather information among the respective weather information allocated to the output layer of the neurocomputer is allocated to the input layer, and the weather information at the arbitrary point L (0) is allocated. Assign the required weather information to the output layer. The contents of the weather information observed at at least one or more different times after an arbitrary time T (t ₀ ) at each point L (j) where the item number is executed, and the arbitrary point assigned by the item number The contents of the weather information observed at least at one or more different times after the arbitrary time T (t ₀ ) of L (0) are input to the neurocomputer, and the respective points L (j ) Learn the state of the change in the atmospheric condition at the arbitrary point L (0) with respect to the change in the atmospheric condition in step. Item No. to the point L (j ₁ ) requiring weather forecast
Run with. Using the learning result of the item number, weather information at each of the points at at least one or more different times before the current time T (t ₁ ) is input to the neurocomputer for each of the points L (j). As a result, necessary weather information after the current time T (t ₁ ) is output, Using learning result of No., for each said each point L (j _1), the neuro-computer inputs the output of No., wherein each point L (j ₁₎ the current time T (t ₁₎ after the Required weather information is output.

As described above, in addition to the item number (1), learning and guessing calculations are performed. In this case, replacement of learning information and assignment of weather information to neurons occur frequently, but are realized with a small-scale neuron configuration. It is possible to enhance the generalization ability of the neurocomputer.

(5) Weather forecasting system III
The weather forecasting system according to any one of claims 2 and 4, and the twelfth, thirteenth, fourteenth, and fifteenth control means according to claim 4) The present system provides a neural computer with respect to any observation point for each weather information. Learning the state of change of only one piece of weather information, further learning the state of change of specific weather information with respect to each change of weather information, and inputting weather information at a different time before the current time, This is for estimating specific weather information at one or more different times thereafter, and will be described with reference to a "neurocomputer basic model in weather forecast system III" shown in FIG.

[0042] At any point, at least at one or more different times, one weather information F (i ₀ ) of the weather information observed and estimated is assigned to each neuron of an input layer and an output layer of a neurocomputer, . At the point, the contents of the one weather information F (i ₀ ) observed before any time T (t ₀ ) and at least one or more different times after the time T (t ₀ ) are stored in a neurocomputer. To learn the state of change of the one weather information F (i ₀ ) at the point, Performing the item numbers with at least one or more of each of the weather information F (i ₀ ) for performing weather observation; The required weather information among the weather information F (i) allocated to the output layer of the neurocomputer in the item number is allocated to the input layer, and one of the weather information F (i ₁ ) selected from the weather information F (i) is selected. ) Is assigned to the output layer, and. At each point, each of the weather information F observed at at least one or more different times after an arbitrary time T (t ₀ )
(I) and the one weather information F observed at at least one or more different times after the time T (t ₀ )
The contents of (i ₁ ) are input to the neurocomputer, and the one weather information F (i ₁ ) corresponding to the change of each weather information F (i)
Learn the state of change. Performing the item numbers ( ₁ ) to ( ₄ ) for each piece of the weather information F (i ₁ ) requiring the weather forecast; Each of the weather information F
For each (i), the current time T (t ₁ ) is stored in the neurocomputer.
By inputting the contents of the one weather information F (i) observed at least at one or more different times before, the one weather information at at least one or more different times after the current time T (t ₁ ) is input. Output F (i); Using the learning result of the item number, each of the weather information F (i ₁ )
A neuro-computer for each, enter the output of No., to output the current time T (t ₁₎ after one weather information F (i _1).

As described above, in addition to the item number (4), the weather forecast of a specific area can be made only by performing the weather observation of that area.
In other words, since prediction can be made independently of other regions, it is not necessary to collect weather information of other regions, and realization can be realized with a small-scale neuro configuration.

(6) Weather forecasting system IV (weather forecasting system according to claim 5 and the sixteenth, seventeenth, eighteenth,
Nineteenth and twentieth control means) The present system checks the chaos (regularity) of time-series data input to a neurocomputer regarding prediction of weather information such as the probability of precipitation and the probability of clearing, and the data is easy to predict. By quantifying whether or not this is the case, the positive factor of the weather information estimated by the neurocomputer (for example,
Sunny, rainy, etc.) or negative (otherwise)
15 and FIG. 1 and FIG.
This will be described using “neurocomputer basic model in weather forecasting system IV” shown in FIG.

[0045] Assigning one of the weather information observed and predicted at at least one or more different times at any point to each neuron for input and output of the neurocomputer; Before any time, and at least one or more different times after the time, the content of the weather information observed at the point, for example, in the case of precipitation probability, during the measurement time, rain is a positive factor If it falls, it is input to the neurocomputer as "1", and if it does not rain as a negative factor, it is input to the neurocomputer as "0", and the output data of the neurocomputer is set to a value from "0" to "1". Also,
The teacher information is “0” or “1” as in the input data.
As one of the above, the state of the change of the weather information at an arbitrary point is learned, and one of the following item numbers is executed.

[0046] (The seventeenth, eighteenth,
Nineteenth control means) (i) using the learning result, inputting the weather information at at least one or more different times before the current time at the point into the neurocomputer,
The weather information at at least one or more different times thereafter is estimated, and by operating the threshold, “1: a positive factor (rain falls)” or “0: negative factor (rain does not fall)” (Ii) Fractal analysis is performed on the set of time-series weather information input to the neurocomputer, and when the regularity is the highest (fractal dimension G = “1”),
Using a mechanism for digitizing the case where the regularity is the lowest (fractal dimension G = “2”) to a value from 100 to 50, the “probability that the neurocomputer guessed the input information” is obtained. The probability of a positive factor or negative factor of weather information, such as the probability of precipitation and the probability of clearing, which is estimated by is calculated from a value between 100% and 50%, and the probability of a positive factor is calculated from 100% to 0%.
Calculate with the value between

[0047] (Seventeenth, eighteenth, and twentieth control means according to claim 5) (i) Using the learning result, a neural computer instructs the neurocomputer to at least one or more different times before the current time at the point. Inputting information, causing the weather information at at least one or more different times thereafter to be inferred, and using this as a firing rate for a positive factor of the weather information; (ii) the time-series weather information input to the neurocomputer; Fractal analysis is performed on the aggregate of, and when the regularity is the highest (fractal dimension G = "1"),
Using a mechanism for digitizing the case where the regularity is the lowest (fractal dimension G = “2”) to a value from 100 to 50, the “probability that the neurocomputer guessed the input information” is obtained. The probability of a positive factor of weather information, such as the probability of precipitation and the probability of clearing, is estimated from a value between 100% and 50%, and (iii) the probabilities of the positive factors of the above-mentioned estimated weather information after the current time. Is multiplied by the "firing rate", and this is output as weather information.

As described above, in addition to the item number (5), the chaos (regularity) of the time series data input to the neurocomputer is examined, and whether or not the data is easy to predict is quantified to estimate the neurocomputer. By obtaining the probability of hitting, it is possible to specify weather information such as the probability of precipitation and the probability of clearing.

[0049]

Embodiments of the present invention will be described below in detail with reference to the drawings.

Embodiment 1 (1) Definition of Weather Information Weather Information: First Weather Information and Second Weather Information First Weather Information: Atmospheric Condition (atmospheric pressure, temperature, humidity,
(Wind direction, wind speed) Second weather information: Weather (precipitation probability, sunny probability, snow probability, heavy rain probability) (2) Basic configuration of information processing device FIG. 1 shows the present invention applied to a weather prediction system of an information processing device. FIG. 2 is a block diagram showing an embodiment of the present invention. (2-1) The information processing apparatus main unit 1 (hereinafter referred to as host 1) includes an arithmetic unit 3 for executing programs and performing data processing, and a host internal bus. 4 and the disk device 5 as a basic configuration.

(2-2) The neuro board 2 assigns the neurons of the input / output layer and the intermediate layer and learns in response to the request through the OS 31 from the weather forecast system application 30 shown in FIG. Initialize information,
Switching of learning information corresponding to seasonal factors, mode setting such as optimization control for learning, data conversion of input / output information of neurocomputer 7, learning of neurocomputer 7, activation for estimation, and neurocomputer 7 In order to download the processing result of the above, a program “neurocomputer control driver 3” described later is used.
(2-2-2) Atmospheric conditions actually observed at each observation point (atmospheric pressure: 860 to 1060 mb, temperature: -50 to 50)
° C, humidity 0-100%, wind direction: 1 for E, S, W, N
Weather information such as 6 divisions, wind speed: 0 to 100 m / s), and weather (precipitation probability, sunny probability, snow probability, heavy rain probability: all%) are converted from '0' to '1' using a sigmoid function.
A programmable hardware operation circuit is incorporated to convert the actual weather information to 2 ⁿ times and convert it into weather information (to be described later), or to convert weather forecast information corresponding to weather information into weather information. Input layer / output layer_corresponding data conversion board 11, (2-2-3). As shown in FIG. 4, the input layer 20 includes an input layer 20, an intermediate layer 21, an output layer 22, and a model input layer (actually, an output layer 22) from which the number of neurons can be set from the outside. The input unit 8 composed of
Is connected to the outside through an output unit 9 composed of a bidirectional latch, and has an input unit 8, an output unit 9, and a control parameter unit 10 for rewriting control parameters of the neurocomputer.
Is actually composed of a single port, and its function can be switched by an external control signal. When the neurocomputer controller 6 inputs weather measurement information (described later) corresponding to weather information actually observed at each observation point to each neuron of the input layer 20 via the neuroboard internal bus 15, Based on the setting or the learning information updated by the neurocomputer 7 itself, weather estimation information (real number obtained by multiplying the value from '0' to '1' by 2 ⁿ : described later) is calculated. Output, and further, teacher information (the input data to the neurocomputer 7) for the output information (estimated value)
Is input to the output layer 22, the error between the output of each neuron of the output layer 22 and the teacher information based on the back propagation rule, and the output of each neuron of the intermediate layer 21 obtained by back-propagating the error. A neurocomputer 7 that learns how the weather condition changes by updating the learning information so as to minimize the error; (2-2-4) an intermediate layer 21 of the neurocomputer 7;
The “coefficients of a plurality of synaptic connections” of each neuron and the “threshold value” corresponding to each neuron in the output layer 22 are composed of an intermediate layer matrix (intermediate layer synaptic coefficient / threshold value) shown in FIG. Matrix), an output layer matrix (a matrix composed of output layer synaptic coefficients / thresholds), a buffer for transferring “learning information” between the learning information current latch 14 and the disk device 5, and a buffer for learning information. A learning information memory 13 which is used as a work area or a temporary save area for intermediate results when learning and estimation calculations are performed in a divided manner; (2-2-5) provided in the neurocomputer 7 and Learning information current latch 14, which currently stores learning information (“threshold value and coefficients of a plurality of synaptic connections” of each neuron), (2-2-6) The neuro-board 2 includes a learning information memory 13 and a learning information current latch 14 or a learning information memory driver 12 for transferring “learning information” between the learning information memory 13 and the disk device 5. It can be inserted into the information processing device main unit 1 by a slot method like a memory board, and as a result, the neurocomputer control device 6 is directly connected to the host internal bus 4.
Alternatively, the arithmetic unit 3 is connected by the host internal bus standard interface 4a, and the neural computer controller 6 or the input layer / output layer_corresponding data conversion board 1
1 has a mechanism capable of writing a command and attached information to an internal memory (not shown) and reading the end report and attached information written in the internal memory in response to an end report from the neurocomputer control device 6;
Learning information (“threshold value and coefficients of a plurality of synaptic connections” of each neuron in the intermediate layer 21 shown in FIG.
When the “threshold value and the coefficients of a plurality of synaptic connections” of each neuron of FIG. In response to the request from the neurocomputer control device 6, the disk device 5 and the learning information memory 13 are driven,
Disk device 5 <-> Neuro computer controller 6
(Through) <-> learning information transfer bus 16 <-> learning information memory driver 12 (through) <-> transfer data through the path of learning information memory 13, In response to a request from the neurocomputer 7, there is a mechanism for transferring data through the path of the learning information current latch 14 <-> learning information memory driver 12 (through) <-> learning information memory 13.

(3) Information processing device software configuration (sixth control means) FIG. 2 is an explanatory diagram showing an example of the function allocation of the information processing device software of the present invention and an interface.

(A) Weather forecasting system application 3
0 (hereinafter, referred to as weather forecast AP30). It is transplanted to the host 1 and sets an observation point for observing weather information and a weather information item to be observed, collects necessary weather information by various means,. The OS 31 is requested to “learn the state of change of the weather condition to the neurocomputer 7 and infer the subsequent weather condition”, and downloads the processing result to the display.
Output to a storage device or a printer.

(B) Neuro computer driver 32
(Control program: hereinafter referred to as OS 31). .. Incorporated in the OS 31 incorporated in the host 1, The request (command issuance) of the weather forecast AP 30 to the neurocomputer 7 is transmitted as it is to the neurocomputer control driver 33, and. The processing result of the neurocomputer 7 is downloaded from the neurocomputer control driver 33, and the weather prediction AP
30 function as a logical interface between the weather forecast AP 30 and the neurocomputer control driver 33,. The disk device 5 and the learning information memory 13 respond to the learning information transfer request of the neurocomputer control driver 33.
Allow data transfer between.

(C) Neurocomputer control driver 3
3 (control program: hereinafter, neuro-computer control driver 33 or neuro-computer control device 6
). Embedded in the neurocomputer controller 6; On the OS 31, the weather prediction AP 30 is provided with a subroutine "individual processing such as learning, estimation, mode setting, etc." for controlling the neurocomputer 7, as a subroutine. The weather information to be learned is assigned to each neuron in the input / output layer, and subsequently managed. Select a learning method, set an arithmetic expression for learning estimation, control parameters, and initialize learning information. The learning information memory driver 12 is controlled to switch to the corresponding learning information according to the seasonal factor which is the unit of learning,. Regarding input / output of the neurocomputer 7, an input layer /
Controlling the output layer_corresponding data conversion board 11 to convert data from the weather information to actual weather information or from weather estimation information to the weather information; In response to the request of the weather forecast AP 30, the neuro computer 7 is requested to perform learning and estimation. The processing result of the neurocomputer 7 is downloaded and released to the OS 31 as it is. Optimization control of learning is performed to increase learning accuracy (reduction of the error between the estimation result and the model) and learning execution speed (reduction of the number of times of learning until the error converges).

(D) Neurocomputer control microprogram 34 (hereinafter referred to as neurocomputer control program 34 or neurocomputer 7) In response to a request from the neurocomputer control driver 33,
By controlling the neurocomputer 7,. Execute the later-described operations for learning and guessing in a straight or divided manner. Set the mode, such as changing arithmetic expressions and control parameters. The learning information is switched according to the seasonal factor which is the unit of learning, and the neuro computer control program 3
4 (neurocomputer 7) executes a series of arithmetic processing shown in FIG. 3 in accordance with a start command of the neurocomputer control driver 33 (neurocomputer control device 6), and stores the arithmetic result in a predetermined location.

The arithmetic expressions used in the inference instruction and the learning instruction will be described in detail in item No. (4-2) “Operations Performed by Learning and Estimation Processes by Neurocomputer”.

(4) Basic Model for Neurocomputer 7 Used in Weather Forecasting System I (4-1) Configuration of Basic Model FIG. 4 shows a basic model of a neurocomputer in the weather forecasting system I. At the observation point at the observation point '', input the time-series weather information before the current time at each observation point, and obtain the weather information at one or more different times after the current time at each observation point. It is a system to make a guess

First, the time T (t ₀ ) (t ₀ :
24, 18, 12, 6 hours before any time k0,
And weather information F _in (t ₀ , j, i ₀ ) for each observation point L (j) corresponding to the current time k0) and the neuron N in the input layer.
_in (t ₀ , j, i ₀ ). In addition, the time T (t ₁ ) (t ₁ ) for performing the weather forecast is 6, 12,
24 hours later), the weather information F _out (t ₁ , j, i ₁ ) for each observation point L (j) corresponding to the output layer neuron N
_out (t ₁ , j, i ₁ ). The optimization of the number of neurons in the input layer 20, the intermediate layer 21, and the output layer 22 will be described later in the item No. (6) “Determining the number of neurons”.

(4-2) Operations Performed by the Neurocomputer 7 in Learning and Estimation Processing The arithmetic operations for learning and estimation performed by the neurocomputer 7 use the following arithmetic expressions [Expressions] to [Expressions]. The above arithmetic expressions are described in sheets 89 and 91 of "Introductory Neurocomputer" (published by 1st edition, 1st edition) by Toyohiko Kikuchi, published by Ohm Co., Ltd. in order to briefly explain the present invention. Quoted arithmetic expressions.

In consideration of the fact that the arithmetic processing from the above equations to the equations can be executed mechanically and the amount to be handled is large, the neurocomputer 7 (a collection of dedicated digital neuron chips or a parallel processing by a multiprocessor) is used. Processing).

[0062] It should be noted that, weather information F _in to be assigned to the input and output layer
(t ₀ , j, i ₀ ) and weather information F _out (t ₁ , j, i ₁ )
Is larger than the number of neurons in the input / output layer,
In particular, the product W _ji X _{i of} “coefficient W _{ji of} synaptic connection” and “input value X _i to neuron j” in the output calculation (formula) of the hidden layer and the output layer, and the calculation of the _hidden layer error (
This is dealt with by calculating the product δ _k W _kj of the “output layer error calculation expression δ _k ” and the “synaptic coupling coefficient W _kj ” in the expression several times.

(A) The dynamic range of the output is set to an arbitrary real number in the [0.1] section by using a “sigmoid function” represented by the following equation as a nonlinear function that determines the input / output characteristics of each neuron. Here, the sigmoid function F (X)
F (X) = (1/2) (1 + tanh (X / U ₀ )) where X: input value U ₀ : slope coefficient of sigmoid function (B) From neurons in the intermediate layer and output layer Output,
The following formula is used for the feedback of the error in the output layer and the feedback of the error in the intermediate layer with respect to the error in the output layer.

[Intermediate Layer and Output Layer Output Calculation] X ′ _j = F (Σ _{i Wji} X _i −H _j ) where X ′ _j : neurons of the intermediate layer 21 and the output layer 22 X _i : input value to neuron j W _ji : synaptic connection coefficient H _j : threshold value of neuron j F (X): approximation of sigmoid function Therefore, the input and output of the neuron are shown in FIG. Become like

[Output Layer Error Calculation] δ _i = (2 / U) (t _i −X _i ) X _i (1−X _i ) where δ _i : error calculation value of output layer i: 1, 2, 3,... N, the number of neurons in the layer X _i : output value of neuron j t _i : model information to neuron j (input data to input layer) U: sigmoid function to neuron for output slope coefficient [intermediate layer error _{calculation] δ 'j = (Σ k} δ k W kj) X j (1-X j) ··· formula where, δ' _j: the intermediate layer error calculated j: 1, 2 , 3... M, the number of neurons in the layer δ _k : calculated value of error in output layer W _kj : coefficient of synaptic connection X _j : output value from target neuron in hidden layer (C) “Atmospheric state” of host 1 learning for the manner of change ", that is the intermediate layer, and the coefficient of synaptic connections of each neuron in the output layer W _ji, and threshold Used below backpropagation rule for modifying the H _j to the optimum value.

W ′ _ji = αW _ji + βδ _j X _j formula H ′ _j = α ′ H _j + β ′ δ _j formula where α, α ′, β, and β ′ are constants W ′ _ji : Modified synaptic coupling coefficient H ' _j : Modified threshold value δ _j : Neuroj output error value X _j : Neuroj output value (Embodiment 2) Next, input / output of neurocomputer 7 Details will be described.

First, an object of the present invention is to make the neurocomputer 7 learn how the weather condition changes, and to estimate the subsequent weather condition based on the previous weather condition. Hereinafter, the weather prediction control will be described.

(5) Management of Weather Information The weather forecast A of the host 1 is obtained from each weather observation point at regular time intervals.
The weather information sent to P30 includes, for example,. First weather information F (i) indicating the atmospheric state: (atmospheric pressure: 860 to 1060 mb: control number i = # 1, temperature:
-50 to 50 ° C: # 2, Humidity 0 to 100%: # 3, Wind speed: 0 to 100 m / s: # 4, above: 0 to 100; Wind direction: East, south, west and north 0 'Northeast North is displayed as '93 .75': # 5). Second weather information F (i) indicating the weather: (Precipitation probability:
# 6, probability of sunny: # 7, probability of snow: # 8, probability of heavy rain: all displayed in%: # 9), and the weather forecast AP 30
As an example, in the past weather information, one year or “season (a collection of weather information having a strong correlation) such as spring, summer, autumn, winter, rainy season” as a unit, Observation time management number (for example, 6 hours is indicated as '1'): (t ₀ ), management number of weather observation point is (j), management number of weather information is (i), each weather observation point Weather information F observed every 6 hours in L (j)
_in (t ₀ , j, i ₀ ), the observation time T (t ₀ ), the observation point L
(J), so that a hierarchy format weather information F _in (i), managed by the disk apparatus 5. The weather information F _in (t
₀ , j, i ₀ ) is a learning and model (teacher) for the neurocomputer 7 to learn “the state of change in weather conditions”.
The latest data is used to estimate future weather information.

The weather forecast AP 30 collects and collects the data
When the learning and estimation of the weather information are requested to the neurocomputer controller 6 using the weather information stored in the as a parameter, the neurocomputer controller 6 controls the input layer / output layer_corresponding data conversion board 11 to convert the weather information. sigmoid function, G (X) = (1/2 ) (1 + tanh ((X-s) / U 1)) ...... ' equation where, G (X): weather Found information _{H in (t 0, j,} i ) X: weather information _{F in (t 0, j,} i) S: G (X) calculated value correction coefficient U _1: real number from using the slope coefficient of the sigmoid function '0''1' and, sigmoidal curve Is set so as to fall within the approximate straight line portion, and is converted into 2 ⁿ times weather measurement information (definition), and the weather measurement information H _in (t ₀ ,
j, i ₀ ). Further, with respect to the operation, by performing the inverse operation, for example, it can be converted weather actual measurement information _{H in (t 0, j,} i 0) weather information _{F in (t 0, j,} i 0) to.

On the other hand, the neurocomputer 7 chronologically sets the weather information H _in (t ₀ , j, i ₀ ) set in each neuron of the input layer 20 (for example, t ₀ = k0-4,
k0-3, k0-2, k0-1, k0 (arbitrary time):
Arbitrary times are set 24, 18, 12, and 6 hours before. Actually, weather conditions after an arbitrary time (k0) are estimated based on a neuron number determination process (described later)). Weather estimation information H _out (t ₁ , j, i ₁ ) (for example, assuming that t ₁ = k0 (arbitrary time), k0 + 1, k0 +
2, k0 + 4: 6, 12, and 24 hours later, but actually outputs [described later] based on the process of determining the number of neurons to each neuron of the output layer 22. This value is higher guess predictive value of neural computer 7, an arbitrary time (k0) after weather actual information _{H in (t 1, j,} i 1) to substantially coincide.

FIG. 5 shows an example of the relationship between neuro computer inputs and outputs in the estimation process.

(6) Neurocomputer Initial Setting (6-1) Processing for Determining the Number of Neurons When the previous weather information is input to the neurocomputer and the subsequent weather information is guessed, "output" is used to increase the predictive accuracy. The number N _{1 of} guess times T (t ₁ ) to be obtained ”(t ₁
Regarding the determination of k0, k0 + 1, K0 + 2, k + 3 or k0, k0 + 1, K0 + 2), “fractal dimension analysis” is introduced. The fractal dimension indicates that the time series data has the same number of dimensions while the tendency is the same, and the larger the number of dimensions, the more complicated the fluctuation of the time series data. In the case of time-series data, the value of the number of dimensions takes a value between “1” and “2”, and the closer to “2”, the more complicated the data is and the more difficult it is to predict. The closer to ', the higher the accuracy of the prediction.

Therefore, when the fractal dimension is close to '1', it is determined that the guess is possible.
The number N ₁ of (t ₁ ) is determined. In addition, “the number N _{0 of} estimated input times T (t ₀ ) to be input” is theoretically “N ₀ = 2 * N _1”.
-1 "is optimal.

For the determination of the intermediate layer, AIC (Akaike's Info
rmation Criterion) to optimize the number of hidden neurons. The definition in neuro-learning is: AIC = (number of learning targets) * log (square error) + 2 * (number of hidden neurons) Equation square error = ΣU _i ((1/2) | V i −W i | ² ) ... where, U: learning target, V: weather measurement information, W: weather estimation information, where AIC value and prediction error have a correlation, and AIC
There is a tendency that the number of hidden neurons that minimize the value of is the same as the number of hidden neurons that minimize the prediction error.

Based on the above background, the weather forecast AP 30
It added measured for each learning object group already weather information _{F in (t 0, j,} i 0) as a parameter: using the "mode setting command input and output layer, and the intermediate layer neurons number of decision"
The input layer 20 in the neurocomputer model is sent to the neurocomputer controller 6 through the OS 31.
A request is made to determine the number of intermediate layers 21 and output layers 22.

In response to the command, the neurocomputer controller 6 calculates the estimated time T (t ₁ ) (t ₁ ) to be output.
The fractal dimension is calculated for each of k0, k0 + 1, K0 + 2, k + 3,..., and the value is close to “1” (do not stay away) “Estimated time T (t ₁ ) number to be output” And repeat in every case,
A statistically optimal estimation time (t
₁ ) Determine the number and determine the number of actual measurement times T (t ₀ ) to be input at the same time.

Therefore, the number of neurons in the input / output layer is as follows.

(Number of input layer neurons) = (number of estimated times t ₀ to be input) * (number of observation points j ₀ ) * (number of weather information i
₀ ) (the number of neurons in the output layer) = (the number of estimated times t ₁ to be output) * (the number of estimated points j ₁ ) * (the number of weather information i ₁ ) actual measurement information _{H in (t 0, j,} i 0), and weather estimation information H _out
With (t ₁ , j, i ₁ ) assigned, the square error (expression) and then the AIC value (expression) are calculated using the number of neurons in the intermediate layer as a parameter, and the intermediate layer with the smallest AIC value is calculated. Is the number of neurons in the intermediate layer in the neurocomputer model used in the weather forecast system.

After the above processing, the neurocomputer controller 6 provides the following information when reporting the end to the weather forecast AP 30 through the OS 31.

Estimated time t to be input t ₀ Estimated time t to be output t ₁ Number of neurons in input layer Number of neurons in output layer Number of neurons in middle layer (6-2) Allocation of weather information to neurocomputer input / output layer Weather forecast AP30 Determines the observation point L (j), the first and second weather information to be observed, and determines the number of neurons in each layer based on the “determination processing of the number of neurons”. 1 OS
31 provided on: using the "mode setting command neurons allocation of the learning object", the actual measurement for weather information F _in (t _0,
j, i ₀ ) and weather information for estimation F _out (t ₁ , j, i ₁ )
To each neuron N _in the input layer of the neurocomputer 7
(t ₀ , j, i ₀ ) and each neuron N _{out of the} output layer
(t ₁ , j, i ₁ ).

In response to the command, the neurocomputer controller 6 sets the input layer neuron assignment management table T10 and the output layer neuron assignment management table T20 shown in FIG.
Requesting that each neuron be set in the input layer 20, the intermediate layer 21, and the output layer 22, and input and output of weather measurement / estimation data to the neurocomputer 7 thereafter are all performed in the input layer neuron assignment management table. T1
0, and output layer neuron assignment management table T20
Manage and control according to

(6-3) Initialization of Learning Information The weather prediction AP 30 uses the “mode setting command: initialization of learning information” to initialize learning information (“synaptic coupling coefficient”, “threshold”). Make settings.

In response to the command, the neurocomputer control device 6 controls the learning information memory driver 12 to write the value of white Gaussian distribution noise as the learning information initial value in the learning information memory 13. If necessary, the learning information transferred to the learning information memory 13 is transferred to the learning information current latch 14.

(7) Learning and guessing of changes in weather conditions (first and second control means according to the first aspect, seventh control means according to the fourth aspect) Control for controlling the neurocomputer 7 to learn how the weather condition changes, and for estimating the weather condition thereafter will be described with reference to the flowcharts shown in FIGS. 8 and 9.

(7-1) Learning of Weather Information (First Control Means According to Claim 1, Seventh Control Means According to Claim 4) The weather forecast AP 30 is configured to change the weather condition of the learning target. In order to make learning, the neuro-computer 7 is used to estimate the subsequent weather conditions from the previous weather conditions, and the learning information is updated so that the error between the estimated weather conditions and the actually observed weather information is minimized. Then, the following “learning command” is issued to the OS 31 (flowchart 1).

[Learning command] + [Control parameter] +
[Learning data] Control parameters: learning target category = year, any of seasonal factors (spring (first half, second half), summer, autumn, winter, rainy season, typhoon, etc.), number of weather information groups of learning data, learning use data: a plurality of weather information set earlier weather information to the item number set _{F in (t 0, j,} i 0) ( training data) (t ₀ = k0-4 as an example, k0-3, k0- 2, k
0-1, k0 (any time); 24,18,12,6 hours ago, at an arbitrary time, weather information) and later to be observed for each observation point corresponding to the input layer 20 weather information F _in (t ₁ , j, i ₁ ) (error correction, learning information update teacher data) (As an example, k0 is an arbitrary time, and t ₁ = k0 + 1,
k0 + 2, k0 + 4; after 6, 12, 24 hours,
(The weather information to be observed at each observation point corresponding to the output layer 22) In response to the command, the OS 31 activates the neurocomputer control device 6 (flow chart 2).

On the other hand, the neuro computer controller 6
Is, weather information F _in for learning _{(t 0, j, i 0} ) and, weather information _{F in (t 1, j,} i 1) the input layer / output layer _ corresponding data of the error correction / learning information for the update By controlling the conversion board 11, the meteorological measurement information H _in (t ₀ , j, i ₀ ) and H _in
(t ₁ , j, i ₁ ) (flowchart 3), and the neuro computer 7 calculates “forecasting the following weather information (item number below) and information actually measured for the estimation” for each of the weather information groups. The update of the learning information (the following item number) is repeated a specified number of times to minimize the error, and the state of change in the weather condition is learned.

[0088] Meteorological information H _in (t ₀ , j,
i ₀₎ is outputted to the input 8 of the neuro-computer 7 starts asked to guess the subsequent weather information to neuro-computer 7 (flow chart 4).

[0089] The neurocomputer 7 responds to the request with the weather estimation information H in accordance with the start command list shown in FIG.
_out (t ₁ , j, i ₁ ) is calculated and stored in the output layer 22, and the end of the estimation processing is reported to the neurocomputer controller 6 (flow chart 5).

. The end report as a trigger, neuro-computer control unit 6 outputs meteorological measured for learning information update information _{H in (t 1, j,} i 1) a (instruction information) to the output portion 9 of the neural computer 7, Neuro Computer 7
And “weather estimation information H _out (t ₁ , j, i ₁ ) calculated by the neurocomputer 7 itself and stored in the output unit 9” and “weather information as the teacher information input to the output layer 22 as a correct answer. actual measurement information _{H in (t 1, j,} i 1) requests that the error "the updating learning information in order to minimize (flow 6).

[0091] In response to the request, the neurocomputer 7 updates the learning information in accordance with the start command list shown in FIG. 3, stores the learning information in the learning information current latch 14, and reports the end of the learning process to the neurocomputer control device 6 (flow chart). 7). The updated learning information is used for the subsequent estimation processing.

[0092] The neurocomputer control device 6
Recognizing that the learning process has been completed on the basis of the end report, the end report is made to the OS 31, and the OS 31 makes a report on the end to the weather forecast AP 30 (flow charts 8, 9).

(7-2) Estimation of weather information (second control means according to claim 1, seventh control means according to claim 4) The weather forecasting AP 30 estimates a weather condition thereafter.
Using the neurocomputer 7, the state of change in the weather conditions of the learning target group is learned, and based on the obtained learning information, the weather conditions actually observed and collected at each observation point are used.
In order to infer the subsequent weather conditions,
Issued to S31 (flowchart 10).

[Guessing command] + [control parameter] +
[Estimation data] Control parameter: learning target category = year, any of seasonal factors (spring (first half, second half), summer, autumn, winter, rainy season, typhoon, etc.) Estimation data: previous weather information F _in (t _0, j, i ₀₎
(Estimation data) (As an example, t ₀ = k0-4, k0-3, k0-2, k
0-1, k0 any time; weather information to be observed at each observation point corresponding to the input layer 20 at any time before 24, 18, 12, 6 hours before) OS31 in response to the command, the neurocomputer The control device 6 is started (flow chart 11).

On the other hand, the neurocomputer controller 6
Is. The estimation data accompanying the command, that is, the weather information F _in (t ₀ , j, i ₀ ) before the time of the last observation,
By controlling the input layer / Output layer _ corresponding data conversion board 11 converts weather actual information _{H in (t 0, j,} i 0) to (flow 12).

[0096] Estimated weather actual measurement information H _in (t ₀ , j, i
₀ ) is output to the input unit 8 of the neurocomputer 7, and the neurocomputer 7 is requested to be activated so as to estimate the weather information after the observation point (flow chart 13).

[0097] The neurocomputer 7 responds to the request with the weather estimation information H in accordance with the start command list shown in FIG.
_out (t ₁ , j, i ₁ ) is calculated and stored in the output unit 9, and the end of the estimation processing is reported to the neurocomputer control device 6 (flow chart 14).

[0098] In response to the end report, the neurocomputer controller 6 causes the weather estimation information H _out (t ₁ , j, i) held in the output unit 9 of the neurocomputer 7.
₁ ) The input layer / output layer_corresponding data conversion board 1
1 to control the estimated weather information F _out (t ₁ , j, i ₁ )
, And store it in the internal buffer (Flowchart 1
5),. The end of the guess command is reported to the OS 31 (flow chart 16).

[0099] On the other hand, the OS 31 occupies the host internal bus 4 at an appropriate timing, and instructs the neuro computer controller 6 to infer the estimated weather information F _out (t ₁ , j) of the input layer / output layer_corresponding data conversion board 11. , I ₁ ) to the memory (flow chart 1).
7),. The neurocomputer control device 6 executes the request and reports an end to the OS 31. The OS 31 reports an end to the weather forecast AP 30, and the estimated weather information F _out (t ₁ , j, i ₁ ). Is set in the memory (flow chart 18).

(Third Embodiment) (8) Management of Learning Information The atmospheric weather condition has a strong time-series factor, and the atmospheric condition generally moves slowly from west to east. Furthermore, due to seasonal factors such as spring, summer, autumn and winter, rainy season, typhoon, and long rain in autumn,
Atmospheric conditions can be roughly classified.

Therefore, for the neurocomputer 7, when learning the state of changes in atmospheric conditions, it is better to limit the range of learning to "every season" rather than "one year" in a time-series manner. Therefore, it is possible to learn a weather state having a strong correlation, and to infer the subsequent weather state in detail.

Therefore, in the present invention, the unit of learning of the neurocomputer 7 is a seasonal factor (spring (first half), spring (second half), rainy season, summer, typhoon, autumn, long rainfall in autumn, winter), and For the factors, weather conditions are learned and estimated, and learning information is managed and controlled for each seasonal factor. The details will be described below.

(8-1) Management of Learning Information (Fourth Control Means of Claim 2) The weather forecasting AP 30 makes the neurocomputer 7 learn the weather forecast, learn the change of the weather condition, estimate the mode, or change the mode. To issue a learning, guessing, or mode setting command to the OS 31 to execute the setting, the corresponding seasonal factor is written in the control parameter area of the command, or a “mode setting command: seasonal factor automatic Judgment (setting / not) "is used to automatically judge the seasonal factor of the weather condition.

On the other hand, the neurocomputer control device 6 manages the learning information shown in FIG. 6 for each seasonal factor by using the seasonal factor type T31 encoding the seasonal factor and the neurocomputer 7 shown in FIG. To indicate whether the learning information can be used, "if the learning information is not created '00'", the learning information is created or initialized, and "if the learning information is stored only on the disc," 01 ". "
"If stored in the learning information memory 13, '8
# '", The usage management flag T32
The disk device 5 for storing the control parameters of the arithmetic expressions for estimation, the storage start address of the learning information memory 13, and the arithmetic expression control parameter storage area T33 indicating the number of bytes to be stored, as well as the learning information (threshold and synapse) The management control is performed using a seasonal factor corresponding learning information management table T30 including a learning information storage area T34 indicating a storage start address of the coefficient (coefficient) and the number of storage bytes.

First, when the learning and estimating process occurs, the neurocomputer control device 6 sets the use status management flag T32 in the seasonal factor corresponding learning information management table T30.
(A) When T32 = '8 #', that is, when the required learning information is stored in the learning information memory 13, it follows the item No. (7) “learning and estimating the state of change in weather conditions”.

(B) T32 = '01 'In the case where learning information is created or initialized and stored only on the disc, the following control is performed.

. The neurocomputer control device 6
A completion report is given to the OS 31 with a status indicating that the learning information has not been set.

. Upon the end report, the OS 31
It requests the neurocomputer controller 6 to transfer the corresponding learning information from the disk device 5 to the learning information memory 13. At the same time, the host internal bus 4 is opened for the neuro board 2.

[0109] In response to the request, the neurocomputer control device 6 refers to the seasonal factor corresponding learning information management table T30 and sends the learning information memory driver 12 a disk with "weather forecasting learning information" corresponding to the relevant seasonal factor. A request is sent from the device 5 to the learning information memory 13.

[0110] After the data transfer is completed, the learning information memory driver 12 reports an end to the neurocomputer control device 6.
31 is reported to the end.

[0111] When the end report is triggered, the OS 31
Requests the neurocomputer controller 6 to perform the learning or the inference process again.

[0112] In response to the request, the neurocomputer control device 6 requests the neurocomputer 7 to perform learning or estimation.

However, when it becomes necessary to transfer the learning information between the learning information memory 13 and the learning information current latch 14, the neurocomputer 7 controls the learning information memory driver 12 to perform the data transfer by itself. .

(C) T32 = '00 ', that is, when learning information is not created, The neurocomputer control device 6 adds a status indicating that no learning information has been created to the OS 31 and the OS 31 and the weather forecast AP 30 and reports the end of the process.

. Weather report AP triggered by the end report
30 executes the “initialization of learning information” shown in the item number (6-3), and then issues a command to the OS 31 again.

Next, when the learning process by the neurocomputer control device 6 is completed, when the completion report is sent to the OS 31, a status indicating the intention to store the updated learning information in the disk device 5 is added. In contrast, OS31
Sends an end report to the weather prediction AP 30 and opens the host internal bus 4 to the neurocomputer control device 6 to permit data transfer of learning information.

(8-2) Judgment of Meteorological Factor (Third Control Means of Claim 1) The weather forecast AP 30 uses the “mode setting command: seasonal factor automatic judgment (setting / not)” to determine the weather. The seasonal factor of the information can be automatically determined.

With the above-mentioned command, with the seasonal factor automatic determination flag set to “set”, the weather forecast AP
30 issues a learning or guessing command, the neurocomputer controller 6 will: Before operating the “weather forecasting system”, the neurocomputer 7 is made to learn weather information, that is, “seasonal factors corresponding to changes in atmospheric conditions”. Using the “seasonal factor determination learning information” created in the item number, determine the seasonal factor of the weather information attached to the command, The following control is performed according to the flowcharts shown in FIG. 11 and FIG. 12 in learning or estimating the state of the change in the atmospheric state using the “weather information for weather prediction” corresponding to the seasonal factor.

(8-2-1) Learning of Seasonal Factors Corresponding to Weather Information (A) The weather forecasting AP 30 sends the following “learning command (seasonal factors)” to the OS 31 to learn the seasonal factors corresponding to weather information. Issued for (Flowchart 2
0).

[Learning command (seasonal factor)] + [Control parameter] + [Learning data] Control parameter: Number of weather information groups of learning data Learning data: Plural weather information groups having a set of item numbers Previous weather information F _in (t ₀ , j, i ₀ ) (learning data) (As an example, t ₀ = k0-4, k0-3, k0-2, k
0-1, k0 (arbitrary time); weather information to be observed at each observation point corresponding to the input layer 20 at an arbitrary time, 24, 18, 12, 6 hours before the weather information group F _in (t ₀ , j, i) corresponding to the seasonal factor (h) (only the seasonal factor (h) corresponding to the weather information is [1], and the other seasonal factors (h) are [0]). The OS 31 activates the neurocomputer control device 6 (flow chart 21).

On the other hand, the neurocomputer controller 6
(C) By setting the input layer neuron assignment management table T10 and the output layer neuron assignment management table T20, the time T (t
_The input weather information F _in (t ₀ , j, i ₀ ) for each observation point L (j) corresponding to the observation point L (j) is converted into the neurons N _in (t ₀ , j, i ₀ ) of the input layer of the neurocomputer 7. ) And the seasonal factors (h) of spring, rainy season, summer, typhoon, autumn, autumn long rain, and winter are assigned to the neurons N _out (seasonal factors) in the output layer, and the “seasonal factor discriminating learning information” is numbered ( Initialize according to 6-3). However, the weather estimation information management number in the output layer neuron allocation management table T20 includes each seasonal factor (h).
Is set (flow chart 22).

(D) The weather information F _in (t ₀ , j, i ₀ ) is controlled by the input layer / output layer_corresponding data conversion board 11 to obtain the actual weather information H _in (t ₀ , j, i ₀ ). (Flowchart 23).

[0123] (E) and the actual seasonal factors (correct) corresponding to "the each the weather information group in neuro-computer 7 meteorological measured information group _{H in (t 0, j,} i 0), neuro-computer 7 is speculated The seasonal factor '(h) (item number below)' is updated a specified number of times to update the seasonal factor discriminating learning information (item number below) to minimize the error. Factor (h) "is learned.

. Outputs weather actual information _{H in (t 0, j,} i 0) to the input 8 of the neuro-computer 7, the neuro-computer 7, so as to estimate the seasonal factors' (h) corresponding to the weather actual information group An activation request is made (flow chart 24).

[0125] The neurocomputer 7 responds to the request according to the seasonal
(H) (Real numbers from [0] to [1]) are calculated and stored in the output layer 22, and the end of the estimation processing is reported to the neurocomputer controller 6 (flow chart 2).
5).

. In response to the end report, the neurocomputer control device 6 determines the seasonal factor (h) (teacher information: the most appropriate seasonal factor for the weather information [1], and the seasonal factor not applicable is [0]) to the neurocomputer. 7 and outputs to the neurocomputer 7, “the neurocomputer 7 itself judges and determines the output layer 22.
In order to minimize the error between the seasonal factor '(h)' held in each neuron and the seasonal factor (h) (the teacher information) input to each neuron in the output layer 22 as a correct answer. A request is made to update the learning information for factor determination (flow chart 26).

[0127] In response to the request, the neurocomputer 7 updates the learning information in accordance with the activation command list shown in FIG. 3, stores the learning information in the learning information current latch 14, and reports the end of the learning process to the neurocomputer control device 6 (flow chart 27). ). The updated “seasonal factor discriminating learning information” is temporarily saved in the disk device 5 and is subsequently used for discriminating the seasonal factor of the weather information.

(F) The neurocomputer controller 6
Recognizing that the learning process has been completed on the occasion of the end report, the end report is made to the OS 31, and the OS 31 makes an end report to the weather forecast AP 30 on this occasion (flow chart 28).

(8-2-2) Determining Seasonal Factors of Weather Information (A) The weather forecast AP 30 learns or predicts the state of change in weather conditions by using the learning command shown in item number (7) or A guess command is issued to the OS 31 (flow chart 30).

(B) In response to the command, the OS 31 activates the neurocomputer controller 6 (flow chart 31).

On the other hand, the neuro computer controller 6
(C) In the input / output layer of the neurocomputer 7, actual measurement weather information F _in (t ₀ , j, i ₀ ) and estimation weather information F _out
When (t ₁ , j, i ₁ ) is assigned, “learning and learning information for estimation” currently used by the neurocomputer 7 and the input layer neuron assignment management table T1
0, the information in the output layer neuron assignment management table T20 is saved in the disk device 5 in the same manner as in the item number (8-1) (flow chart 32).

(D) The “seasonal factor discrimination learning information” created in the item number (8-2-1) is read from the disk device 5 in the same manner as the item number (8-1). Transfer to latch 14. Furthermore, the input layer neurons allocation management table T10, an output layer neurons allocation management table T20, and according Koban (6-2), found for weather information _{F in (t 0, j,} i 0), and the weather information group Is assigned to each neuron N _in (t ₀ , j,
i ₀ ) and neuron N _out (seasonal factor) (flow chart 33).

(E) Weather information F attached to the command
To determine the seasonal factor of _in (t ₀ , j, i ₀ ),. By controlling the input layer / Output layer _ corresponding data conversion board 11 converts weather actual information _{H in (t 0, j,} i 0) to (flow 34).

[0134] The weather measurement information H _in (t ₀ , j, i ₀ ) is output to the input unit 8 of the neurocomputer 7, and the neurocomputer 7 receives the seasonal factor corresponding to the weather measurement information.
A start request is made to estimate (h) (flow chart 35).

[0135] In response to the request, the neurocomputer 7 uses the “seasonal factor discriminating learning information” and according to the start command list shown in FIG.
The real number up to [1] is calculated and stored in each neuron N _out (seasonal factor) of the output layer 22 and the end of the estimation processing is reported to the neurocomputer controller 6 (flow chart 36).

[0136] The neurocomputer control device 6
The neurocomputer 7 is connected to each neuron N in the output layer 22.
_out (seasonal factor), the neuron N that outputs the largest value among the seasonal factor values for the weather information
_out (seasonal factor) is determined as a seasonal factor for the weather information (flow chart 3).
7).

(F) (Fifth control means according to claim 2)
The neurocomputer control device 6 operates in the same manner as in item (8-1). "Learning information for learning estimation" saved in the disk device 5
Is transferred to the learning information current latch 14 of the neurocomputer 7, and. The input layer neuron assignment management table T10 and the output layer neuron assignment management table T20 are transferred to the memory of the neurocomputer controller 6, and further, the input layer neuron assignment management table T10, the output layer neuron assignment management table T20, and the item number (6
In accordance with -2), the actual weather information F _in (t ₀ , j, i ₀ ) and the estimation weather information F _out (t ₁ , j, i ₁ ) are transmitted to each neuron N in the input / output layer of the neurocomputer 7. _in (t ₀ ,
j, i ₀ ) and N _out (t ₁ , j, i ₁ ) (flowchart 38).

. According to the item number (7), the state of change of the weather condition is learned and estimated (flow chart 39).

The invention made by the present inventor has been specifically described based on the embodiments. However, the present invention is not limited to the above embodiments, and various modifications can be made without departing from the gist of the invention. Needless to say.

For example, in the above embodiment of the present invention, as an example, weather forecasting (learning and estimating the state of a change in atmospheric state with a strong time series factor) has been described. Basic configuration diagram of processing equipment ",
If "the logical interface of the AP, OS, and neuro control of FIG. 2" is used, the application field is considerably wide.

[0141]

Advantageous effects obtained by typical ones of the inventions disclosed in the present application will be briefly described.
It is as follows.

According to the weather forecasting system and the neurocomputer control system of the present invention, (1) the neurocomputer learns weather observations and the state of atmospheric changes at a plurality of observation points where weather forecasting is performed. In the neurocomputer, previously (eg, 24, 1
By inputting the weather information 2,6 hours before and at the present time, it is possible to infer the weather information after that (for example, 6,12,24 hours later).

[0143] Since observation and prediction of weather information can be set arbitrarily, a nationwide version of weather information is provided.
In addition to predicting the nationwide version of the weather, local information of any area can be predicted by adding the weather information of the area where the weather is to be predicted to the learning target of the neurocomputer.

[0144] The operation of the weather forecasting system according to the present invention does not require specialized knowledge on weather.

(2) “Seasonal factors” corresponding to “weather information” (eg, spring, rainy season, summer, typhoon, autumn, long rain in autumn, winter)
Is learned and guessed (determined) in the same manner as in item number (1).
By doing so, the range of learning for the weather of the neurocomputer can be limited to "every season", the weather conditions with strong correlation in time series can be learned, the subsequent weather conditions can be determined in detail, and especially at the turn of the season The forecast system can be maintained.

(3) The information processing apparatus connects the “neurocomputer controller connected to the neurocomputer” by an internal bus or a standard interface, and executes a subroutine provided by the neurocomputer control system on the OS by command means. By calling from the application program, the processing using the neurocomputer can be executed.

(4) The neurocomputer learns the state of the change of the atmospheric state independently at each point where the weather observation is performed, and further, the state of the change of the atmospheric state at a specific point with respect to the change of the atmospheric state at each point. By learning the weather information at different times before the current time at each observation point and estimating the weather information at one or more different times after the current time at a specific point, this system It can be realized with a neuron configuration, and the generalization ability of the neurocomputer can be enhanced.

(5) At an arbitrary observation point, the neural computer is made to learn the state of change of only one piece of weather information for each piece of weather information. By learning the situation, inputting weather information at different times before the current time, and inferring specific weather information at one or more different times after the current time, Simply by performing weather observations in a specific area, it is possible to forecast the weather in that area, that is, it is possible to forecast independently of other areas, and it is not necessary to collect weather information in other areas.

[0149] This system can be realized with a small-scale neuron configuration, and the generalization ability of the neurocomputer can be enhanced.

(6) Regarding the prediction of the second weather information such as the probability of precipitation and the probability of clearing, the chaos (regularity) of the time-series data input to the neurocomputer is examined to determine whether the data is easy to predict. From the “probability that the neurocomputer guesses the input information” obtained using a mechanism for digitizing, the positive factors of the second weather information guessed by the neurocomputer (for example, sunny, rainy, etc.) , Or by determining the probability of hitting against a negative factor (otherwise), it is possible to specify second weather information such as the probability of precipitation and the probability of clearing.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating an example of an information processing apparatus using a neurocomputer.

FIG. 2 is an explanatory diagram illustrating an example of an interface for neuro-control of an AP-OS.

FIG. 3 is an explanatory diagram showing an example of a list of startup instructions to a neurocomputer.

FIG. 4 is an explanatory diagram showing an example of a neurocomputer basic model in the weather forecasting system I.

FIG. 5 is an explanatory diagram showing an example of the relationship between neurocomputer input and output.

FIG. 6 is an explanatory diagram showing an example of neurocomputer learning information and input / output of neurons.

FIG. 7 is an explanatory diagram showing an example of a neuron assignment management table.

FIG. 8 is a flowchart illustrating an example of control of a learning process by a weather prediction AP.

FIG. 9 is a flowchart illustrating an example of control of estimation processing by a weather forecast AP.

FIG. 10 is an explanatory diagram showing an example of a seasonal factor corresponding learning information management table.

FIG. 11 is a flowchart illustrating an example of control of a learning process of a seasonal factor corresponding to weather information.

FIG. 12 is a flowchart illustrating an example of control of a process of estimating a seasonal factor corresponding to weather information.

FIG. 13 is an explanatory diagram illustrating an example of a neurocomputer basic model in the weather forecast system II.

FIG. 14 is an explanatory diagram showing an example of a neurocomputer basic model in the weather forecast system III.

FIG. 15 is an explanatory diagram showing an example of a neurocomputer basic model in the weather forecast system IV.

FIG. 16 is an explanatory diagram illustrating an example of prediction_post-processing of “second weather information”.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Information processing device main body 2 Neuro board 3 Arithmetic unit 4 Host internal bus 4a Host internal bus standard interface 5 Disk device 6 Neuro computer control device 7 Neuro computer 8 Input unit 9 Output unit 10 Control parameter unit 11 Input layer / output layer_ Corresponding data conversion board 12 Learning information memory driver 13 Learning information memory 14 Learning information current latch 15 Neuroboard internal bus 16 Learning information transfer bus 20 Input layer 21 Middle layer 22 Output layer 30 Weather forecasting system application (AP) 31 Operating system (OS) ) 32 Neurocomputer driver 33 Neurocomputer control driver 34 Neurocomputer control microprogram

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-2-221888 (JP, A) JP-A-4-58188 (JP, A) JP-A-5-81228 (JP, A) (58) Field (Int.Cl. ⁷ , DB name) G01W 1/10 G01W 1/02 G06F 15/18

Claims (5)

(57) [Claims] [Claim 1] to collect <br/> shown to meteorological information the atmospheric conditions of a plurality of points to carry out meteorological observations in more than one time, out of the previous SL a plurality of points, one or more of one place or more points A weather forecasting system for estimating weather information of a plurality of locations for performing weather observations collected at a plurality of first times
Weather information indicating the atmospheric conditions of the
A first control means for powering and a neurocomputer based on said weather information
Using the learning information, the second time after the first time
Weather information of one or more of the plurality of points is estimated.
A second control means for measuring and outputting, comparing the estimated value at the second time with the observed value at the second time.
And a third control means for updating the learning information . 2. The weather forecast system according to claim 1, wherein
A fourth control means for identifying a plurality of similar patterns;
At each pattern, the neurocomputer changes atmospheric conditions.
Fifth control means for learning the state of formation and creating learning information
With the door, the second control means, learning information corresponding to the pattern
The weather forecasting system according to claim 1, wherein weather information is estimated using the information . 3. An information processing apparatus comprising: an arithmetic unit for executing a weather forecasting program and performing data processing; a memory for storing the weather forecasting program; and an internal bus connecting the arithmetic unit and the memory. And a neuro-board connected to an internal bus of the information processing device and having a neuro-computer control device for controlling the neuro-computer, wherein the neuro-computer control device performs weather observation Receiving weather information collected at a plurality of points at a plurality of times from an information processing apparatus executing the weather forecast program , and learning information based on the weather information.
The plurality of the plurality of times at a second time after the used first time
Estimating the weather information of one or more of the points,
To compare the estimated value at the second time with the observed value at the second time
A neuro-computer control system, which controls updating of the learning information by the user . 4. A method of allocating necessary weather information out of the weather information estimated and estimated at a plurality of points at a plurality of times to each neuron for input and output of the neurocomputer. The contents of the weather information observed at a plurality of points at a plurality of times before any time and at one or more points at one or more times after the time are input to the neurocomputer. A second step of performing, based on the content of the weather information before the arbitrary time, the content of the weather information after the arbitrary time estimated by the neurocomputer, and the content of the weather information observed after the arbitrary time. An operation of updating the learning information (coupling weight, threshold value) used by the neurocomputer by feeding back an error from the content of the weather information is performed when the error is sufficiently small. A third step of repeating until using the learning information, the neuro-computer,
A fourth step of inputting the weather information of each point at a plurality of times before the current time to output necessary weather information of each point after the current time, and a seventh step of performing a control operation comprising: A first function having control means, learning, by the neurocomputer, a state of a change in an atmospheric state at the point alone for each of at least one or more points for observing necessary weather information among the weather information; Eighth control means for causing the neurocomputer to learn a state of a change in the atmospheric state at an arbitrary point selected from the respective points corresponding to the change in the atmospheric state at the respective points. Means, using the learning result of the eighth control means, for each of the points, input the weather information at a plurality of times before the current time of the point to the neurocomputer, and Few At least one of the tenth control means for outputting the weather information necessary for at least one different time, and using the learning result of the ninth control means, for each arbitrary point selected by the ninth control means. Inputting the output result of the tenth control means to the neurocomputer;
An eleventh control means for outputting the required weather information at least one or more different times after the current time; a second function comprising: observing necessary weather information among the weather information to the neurocomputer; A twelfth control unit that independently learns the state of change of the weather information at an arbitrary point for performing one weather information, and the neurocomputer corresponds to the change of each weather information at the arbitrary point. A thirteenth control unit that learns a state of change of one weather information selected from the weather information at the arbitrary point; and a learning result of the twelfth control unit, for each of the weather information. Inputting the one weather information of at least one or more different times before the current time of the arbitrary point, and outputting the one weather information of at least one or more different times after the current time. Using a learning result of the fourteenth control means and the thirteenth control means,
For each arbitrary weather information selected by the control means, the output result of the fourteenth control means is input to the neurocomputer, and at least one or more different times after the current time of the arbitrary point First to output one weather information
The weather forecasting system according to claim 1 or 2, or the neurocomputer control system according to claim 3, which has at least one of the following control functions: 5. The neurocomputer displays, at an arbitrary point for performing weather observation, a state of a change in one of weather information such as a probability of precipitation, a probability of clearing, a probability of snow, a probability of heavy rain, and the like. Sixteenth control means for learning, and using the learning result of the sixteenth control means, inputting the one weather information at a plurality of times before the current time at the point to the neurocomputer, and after the current time A seventeenth control unit that outputs the one weather information at one or more times, and a chaos property of the input weather information, and estimates a probability of hitting the weather information output by the seventeenth control unit. A fourth function comprising: eighteenth control means. The weather information output by the neurocomputer in the eighteenth control means is set to “1” by operating a threshold value.
Or converting it to “0”, separating the weather information into positive factors and negative factors, and using a mechanism to digitize the case where the regularity of the data is the highest from the case where the regularity is the lowest.
The fractal dimension of the temporal change of the input weather information is checked, and the probability that the estimated weather information has a positive factor or a negative factor is determined as a value between 100% and 50%, and the positive factor or the negative of the weather information is determined. The probability that the factor hits is 10
A fifth function comprising a nineteenth control means and the sixteenth and seventeenth control means, which are determined by a value between 0% and 0%. The eighteenth control means outputs the weather information output by the neurocomputer. , The firing rate for the positive factors of the weather information, from the highest regularity of the data,
The fractal dimension of the temporal change of the input weather information is examined by using a mechanism for quantifying the case where the regularity is the lowest, and the probability of the positive factor of the estimated weather information being 100 is determined.
% And a twentieth control means for obtaining the product of the firing rate and the probability of being a positive factor of the estimated weather information as weather information, and the sixteenth and seventeenth control means. The weather forecasting system according to claim 1 or 2, further comprising at least one of a sixth function and a neuro-computer control system according to claim 3.