JPH10239452A - Rainfall and snowfall forecasting device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To ensure an accurate forecasting even in the case of generation of new echo pattern and the hard movement thereof by evaluating the accuracy of a forecasting value generated by a neural circuit network and the accuracy of a forecasting value generated by use of moving vector from a radar image, and then determining a final forecasting value. SOLUTION: A neural circuit network model forecasting mechanism 101 sends an obtained forecasting value 301 to an output control mechanism 103. A mutual correlation model forecasting mechanism 102 obtains a moving vector making a mutual correlation value, from two sheets of echo patterns at an arbitral time interval from a meteological radar. A forecasting value obtained from the moving vector and a vector 302 are fed to the output control mechanism 103. The mechanism 103 performs comparison between the forecasting value 301 and vector 302 and an actually measured image, and an elapse of time where the forecasting accuracy of the former is made superior to that of the latter is obtained. The passing time is used as a threshold, and the forecasting value of the forecasting mechanism 102 is adopted as a final forecast image until the time passes from the start of forecasting, and thereafter the forecasting value of the forecasting mechanism 101 is adopted as the final forecast image.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a weather forecasting apparatus having a neural network model for learning weather dynamics from a measured weather radar image.

[0002]

2. Description of the Related Art Conventionally, a measured radar image is given to a neural network model to learn the movement of an echo pattern representing a rainfall and snowfall area, and a weather radar image is predicted using the learned neural network model. (For example, Japanese Patent Application Laid-Open No. 7-20255, "Parallel calculation type weather radar image prediction device", and Japanese Patent Application Laid-Open No.
No. 63861, “Parallel calculation type weather radar image prediction device”, and JP-A-7-128456 “Nonlinear parallel calculation type weather radar image prediction device”).

In these apparatuses, a measured radar image is given to a neural network model having a product-sum calculation unit to train weather dynamics, and a weather radar image is trained using the trained neural network model. By predicting the amount of rainfall and snowfall after a certain time at each grid point of the above, and repeatedly predicting the amount of rainfall and snowfall after a certain time using the neural network model using this prediction value as input Up to the predicted value.

In these apparatuses, it is necessary to provide a weather radar image for a certain period of time to a neural network model in order to learn weather dynamics, and sufficient weather radar image data such as when rainfall and snowfall starts is obtained. At a stage where it is not performed, there is a disadvantage that sufficient prediction accuracy cannot be obtained because learning of the neural network model is not sufficiently performed. In addition, when the moving speed of the echo pattern in the weather radar image is high, the prediction process is repeated again with the prediction result obtained using the neural network model as input, so that sufficient accuracy can be obtained when the prediction time becomes long. There is a disadvantage that it cannot be done. Japanese Patent Laid-Open Publication No. Hei 8-50181, entitled "Rainfall Movement Prediction Apparatus," has a concept of estimating a movement vector from weather radar images by "a plurality of methods" and obtaining an "optimum movement vector" based on the result. A rainfall prediction device has been released.

[0005] The calculation of the optimum movement vector is described as follows. "When obtaining the current optimum vector based on a plurality of movement vectors, each movement vector is decomposed in two orthogonal directions.
A direction component is obtained, and an optimum movement vector is obtained from the largest component among the movement vectors. "

[0006]

SUMMARY OF THE INVENTION The object of the present invention can be applied to a rainfall / snowfall prediction device using a weather radar image as described above, even when a new echo pattern is generated or when the echo pattern moves rapidly. Another object of the present invention is to provide a rainfall / snowfall prediction device with good prediction accuracy without using the above-described component decomposition of the movement vector and the histogram (calculation of the most common component).

[0007]

A rainfall and snowfall prediction apparatus according to the present invention is provided with a weather radar image to a neural network model to learn weather dynamics, and uses the trained neural network model to learn the amount of rainfall and snowfall. In a device that performs short-term prediction of, a plurality of sets of two weather radar images measured at an arbitrary time interval are trained by a neural network model, and a short-time From the first prediction means for predicting later rainfall and snowfall and two weather radar images measured at an arbitrary time interval, a movement vector of an echo pattern in the weather radar image is calculated, and from this movement vector, A predicted value of rainfall snowfall information obtained from a second prediction means for predicting a rainfall snowfall amount after a short time;
By comparing and evaluating the measured value of the rainfall and snowfall information obtained from the weather radar, the elapsed time and the movement vector since the start of the prediction,
The present invention is characterized in that there is provided a third predicting means for performing output control as a predicted value of the predicted rainfall and snowfall obtained from the predicting means.

[0008] The third predicting means inputs a radar image to be predicted to the first predicting means and the second predicting means, obtains each predicted value, and calculates a predicted value of the radar image. Evaluating means for comparing prediction accuracy and performing processing for obtaining an elapsed time from the start of prediction until the prediction accuracy of the first prediction means becomes better than the prediction accuracy of the second prediction means for a large amount of radar images. The elapsed time obtained by the evaluation means as a threshold, and from the start of prediction until the threshold is reached, the prediction value of the second prediction means is output as the predicted rainfall and snowfall, And control means for outputting a predicted value by the first predicting means as predicted rainfall and snowfall after the time exceeding.

Further, the third predicting means inputs a radar image to be predicted to the first predicting means and the second predicting means, obtains each predicted value, and calculates a predicted value of the radar image. An editing unit that compares the prediction accuracy and edits the prediction value of the prediction unit with a better prediction accuracy into an application area table; and, based on the application area table, calculates the prediction value of the prediction unit along with the elapsed time based on the predicted rainfall and snowfall. Control means for outputting as

[0010] Further, the third predicting means stores, for the first predicting means and the second predicting means, respective predicted values obtained by inputting a radar image to be predicted for a predetermined time. Means, calculating means for calculating the CSI value at each time using the actually measured value obtained after the lapse of the predetermined time, and interpolation polynomial of degree N-1 given by Lagrange's equation passing through N points. The CSI value after an arbitrary time is predicted by using an equation means for obtaining an interpolation polynomial of degree N-1 passing through the prediction time and each CSI value at each time with respect to the prediction value, and using the interpolation polynomial. Interpolation means for outputting a predicted value as an interpolated value based on the prediction result may be included.

Further, the third predicting means may control only the necessary predicting means by starting and stopping the first predicting means and the second predicting means.

[0012] Furthermore, the prediction accuracy may be such that “the area of both rainfall” and “the area of rainfall” correspond to the corresponding area of the actually measured image and the corresponding area of the predicted image.
A CSI value indicating a hit rate defined by a ratio to the sum of “the regions both raining” and “the two opposite regions of the actual measurement and the prediction rainfall” may be used.

Further, the third predicting means stores, for the first predicting means and the second predicting means, respective predicted values obtained by inputting a radar image to be predicted for a predetermined time. Means, calculating means for calculating the CSI value at each time using the actually measured value obtained after the lapse of the predetermined time, and interpolation polynomial of degree N-1 given by Lagrange's equation passing through N points. The CSI value after an arbitrary time is predicted by using an equation means for obtaining an interpolation polynomial of degree N-1 passing through the prediction time and each CSI value at each time with respect to the prediction value, and using the interpolation polynomial. An interpolating means for outputting a predicted value as an interpolated value according to a prediction result; and, if the predicted value using the interpolating polynomial is higher than the predicted value of the second predicting means, the predicted value is predicted rainfall and snowfall. As quantity Or the third prediction means including control means for force.

[0014]

Embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a configuration diagram of an embodiment of the present prediction device. This device includes a neural network model prediction mechanism 101,
It comprises a cross-correlation model prediction mechanism 102 and an output control mechanism 103.

The neural network model predicting mechanism 101 learns a pair of echo patterns at a certain time T and T + h a plurality of times by using a neural network model for an echo pattern from a weather radar arriving at regular time intervals (h). Then, using the neural network model after the learning is completed, a prediction value for a certain time (h) is obtained, and a prediction for a certain time (h, 2h ahead in total) is performed again using the obtained prediction value as an input. By repeating the operation, a predicted value at an arbitrary time point is obtained. The neural network model operating in this mechanism constitutes a layered network having one input layer, a plurality of intermediate layers, and one output layer, and each layer is composed of units, weights, and bias values. The unit obtains the product sum of the output value (N) of the unit of the previous layer and the weight of the unit, and outputs a value obtained by adding the bias value. Various methods can be applied to the input layer, the intermediate layer, the configuration of the output layer, and the method of calculating the product sum of units.
It is not limited by the present invention. This mechanism transfers the obtained predicted value 301 to the output control mechanism 103.

The cross-correlation model predicting mechanism 102 calculates the echo pattern from the weather radar arriving at regular time intervals (h) from two echo patterns at an arbitrary time interval (ΔT). According to the method shown, a cross-correlation value is calculated, and a movement vector that maximizes the cross-correlation value is obtained. Using this movement vector, a predicted value after an arbitrary time is obtained. When calculating the cross-correlation value, a method of calculating the entire area covered by the weather radar as one, or a method of dividing the entire area into arbitrary multiple areas and calculating the movement vector for each area and combining them is applicable The present invention can be combined with any linear prediction method. This mechanism transfers the obtained predicted value and the movement vector 302 to the output control mechanism 103. "Reference 1: Yoshio Yuma, Katsuhiro Kikuchi, Imahisa," On the moving speed of echoes by a simple weather radar ", Hokkaido University Geophysical Research Report, 44, pp. 23-34." 201, neural network model prediction mechanism 101, cross-correlation model prediction mechanism 102
Using the respective prediction results 301 and 302 transferred from the
It is determined by any of the following methods in consideration of available computer power and the like.

The following method is called a “neural circuit model”.
By comparing the movement vector of the rainfall and snowfall obtained from the transition of the prediction value predicted by the above with the movement vector of the rainfall and snowfall predicted by “arbitrary means”, it is possible to select only one of the prediction values. It is a technique that predicts the amount of rainfall and snowfall based on the technology of performing a combination process such as weight average of both, threshold value processing, etc., and it is known that `` when finding the current optimal vector based on a plurality of Decomposition into two orthogonal directions to obtain the two-direction components, and to determine the optimum motion vector from the largest component of each motion vector ", and use a histogram (calculation of the most component). Not a way.

Method 1: Neural Network Model Prediction Mechanism 101
And the cross-correlation model prediction mechanism 102, the radar image to be predicted is input, each predicted value is obtained, and the measured value is compared with the actually measured image, and the prediction accuracy of the former is better than the prediction accuracy of the latter The processing to find the elapsed time from the start of prediction up to is evaluated for a large number of radar images. The elapsed time obtained as a result is set as a threshold value, and the cross-correlation model prediction mechanism 102
Is used as the final predicted image, and then the predicted value of the neural network model prediction mechanism 101 is used as the final predicted image.

Further, the output control mechanism 103 controls the start and stop of the neural network model prediction mechanism 101 and the cross-correlation model prediction mechanism 102 using the control information 303, so that only the necessary prediction mechanism operates. To control.

The prediction accuracy is compared with the CSI (Critical Success Index) defined below.
x), but it is also possible to use any evaluation criterion such as the sum of squared errors.

[0021]

(Equation 1) Here, N11: a measured image is a rainfall region and a predicted image is also a rainfall region N10: a region where the measured image is not a rainfall region, and a predicted image is a rainfall region N01: a region where the measured image is a rainfall region and the predicted image is not a rainfall region Method 2 In the same manner as described above, the motion vector obtained from the cross-correlation model prediction mechanism 102 and the neural network model predictor 10 are obtained by the preliminary analysis using the radar image to be predicted.
By comparing 1 with the predicted value of the cross-correlation model prediction mechanism 102 and the actual measurement image, a motion vector and an application area table of the two prediction mechanisms are obtained, and a final predicted image 202 is obtained based on the application area table.

Further, the output control mechanism 103 controls the start and stop of the neural network model prediction mechanism 101 by using the control information 303, so that the prediction calculation load can be reduced.

The prediction accuracy is compared with the CSI as described above.
, But other evaluation functions can be used in the same manner as in the method 1.

Method 3: Neural Network Model Prediction Mechanism 101
And a predicted value 30 obtained by inputting a radar image to be predicted to the cross-correlation model prediction mechanism 102.
1 and 302 are accumulated for a fixed time (x ₁ , x ₂ ,..., X _n ). Next, the above-mentioned CSI value is calculated for each time using the actually measured value obtained after the lapse of this time.

N points y ₁ = F (x ₁ ), y ₂ = F (x
₂ ),... It is known that an interpolation polynomial of degree N−1 passing through y _N = F (x _n ) is given by the following Lagrange equation.

[0026]

(Equation 2) According to the above equation, the predicted times X ₁ , X ₂ ,... X _n and the CSI values CSI ₁ , CSI ₂ ,. . . CS
The interpolating polynomial of order N-1 through I _n, obtains the prediction value 301 and the predicted value 302. Using this interpolation polynomial, the CSI value after an arbitrary time is predicted, and based on the prediction result, the output of the prediction mechanism that can be expected to output better accuracy is controlled to be the final prediction result. Until the evaluation result using the interpolation polynomial is obtained, the predicted value of the cross-correlation model prediction mechanism 102 is unconditionally set as the final predicted image 202.

Although the present embodiment has been described using the neural network model predicting mechanism and the cross-correlation model predicting mechanism, output control is performed by using other predicting mechanisms together. In addition to selectively outputting predicted values to be output, output control such as weighted averaging can be easily extended.

[0028]

In the rainfall / snowfall prediction apparatus using the conventional neural network model, the neural network model does not complete the learning of the actual radar image, such as at the start of the rainfall, or the echo in the measured radar image. If the movement of the pattern is fast, there is a drawback that accurate prediction results cannot be obtained.

According to the present invention, prediction means using a neural network model, and means for predicting using a movement vector obtained by calculating a cross-correlation value from radar images obtained at arbitrary time intervals, By the third means for evaluating the accuracy of the plurality of predicted values and controlling the output of the final predicted value, an accurate predicted result can be obtained even when the neural network model cannot obtain sufficient accuracy.

[Brief description of the drawings]

FIG. 1 is a block diagram of one embodiment of a rainfall / snowfall prediction device of the present invention.

[Explanation of symbols]

Reference Signs List 101 Neural network model prediction mechanism 102 Cross-correlation model prediction mechanism 103 Selection output mechanism 201 Actual measurement image (input data) 202 Predicted image (output data) 301 Predicted value output from neural network model prediction mechanism 302 Cross-correlation model prediction mechanism Predicted value to be output and movement vector 303 Control information for controlling the prediction mechanism by the output control mechanism

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Koichi Watanabe 223-1 Yamashita-cho, Naka-ku, Yokohama-shi, Kanagawa Prefecture Inside NTT Software Corporation (72) Inventor Masahiro Tokioka Yamashita, Naka-ku, Yokohama-shi, Kanagawa Prefecture 223-1 Cho, NTT Software Corporation

Claims (7)

[Claims] An apparatus for learning weather dynamics by giving a weather radar image to a neural network model and performing a short-term prediction of rainfall and snowfall using the trained neural network model, comprising: Prediction means for making a neural network model learn a plurality of sets of two meteorological radar images measured at a time, and using the neural network model after learning to predict rainfall and snowfall shortly afterward. And calculating a movement vector of an echo pattern in the weather radar image from the two weather radar images measured at an arbitrary time interval, and predicting the amount of rainfall and snowfall after a short time from the movement vector.
The predicted value of the rainfall and snowfall information obtained from the prediction means, the measured value of the rainfall and snowfall information obtained from the weather radar, and the elapsed time and the movement vector after the start of the prediction are compared and evaluated. A rainfall and snowfall prediction device, comprising: third prediction means for performing output control as a predicted value of rainfall and snowfall. 2. The third predicting means inputs a radar image to be predicted to the first predicting means and the second predicting means, and obtains respective predicted values,
A comparison of the prediction accuracy with the measured image is performed, and the processing for obtaining the elapsed time from the start of the prediction until the prediction accuracy of the first prediction unit becomes better than the prediction accuracy of the second prediction unit is performed on a large amount of radar images. Evaluation means to perform, the elapsed time obtained by the evaluation means as a threshold,
From the start of prediction until the threshold is reached, the predicted value of the second predictor is output as the predicted rainfall snowfall, and after the time exceeding the threshold, the predicted value of the first predictor is calculated as the predicted rainfall. Third including control means for outputting as a quantity
2. The rainfall and snowfall prediction device according to claim 1, further comprising: 3. The third predicting means inputs a radar image to be predicted to the first predicting means and the second predicting means, and calculates respective predicted values,
Comparing the prediction accuracy with the actually measured image, editing means for editing the prediction value of the prediction means with better prediction accuracy into the applicable area table, and, based on the applicable area table, the predicted value of the corresponding prediction means together with the elapsed time. The rainfall and snowfall prediction device according to claim 1, further comprising third prediction means including control means for outputting the predicted rainfall amount. 4. The third predicting unit has a third predicting unit that controls only the necessary predicting unit by controlling the start and stop of the first predicting unit and the second predicting unit. The rainfall and snowfall prediction device according to claim 1. 5. The prediction accuracy is as follows: For regions of the actually measured image and the predicted image corresponding to each other, “both regions of rainfall” and “two regions of rainfall both of actual measurement and prediction” of “region of both rainfall” 4. The rainfall and snowfall prediction device according to claim 2, wherein the prediction accuracy is determined using a CSI value indicating a hit ratio defined as a ratio with respect to a sum of the rainfall and snowfall. 6. The storage device according to claim 3, wherein the third predictor stores, for the first predictor and the second predictor, a predicted value obtained by inputting a radar image to be predicted for a predetermined time. Means, using an actual measurement value obtained after the lapse of the predetermined time,
Calculating means for calculating an I value for each time; and an interpolation polynomial of an order N-1 given by Lagrange's equation passing through N points, the prediction time and the order N-1 passing through the respective CSI values at each time. Formula means for obtaining an interpolation polynomial for the predicted value; interpolating means for predicting a CSI value after an arbitrary time using the interpolation polynomial, and outputting a predicted value as an interpolation value based on the prediction result; A third predicting unit including a control unit that outputs the predicted value as a predicted rainfall amount if the predicted value using the polynomial is higher than the predicted value of the second predicting unit and the prediction accuracy is high. 1. The rainfall and snowfall prediction device according to 1. 7. The storage device according to claim 3, wherein the third prediction unit stores, for a certain period of time, a prediction value obtained by inputting a radar image to be predicted for the first prediction unit and the second prediction unit. Means, using an actual measurement value obtained after the lapse of the predetermined time,
Calculating means for calculating an I value for each time; and an interpolation polynomial of an order N-1 given by Lagrange's equation passing through N points, the prediction time and the order N-1 passing through the respective CSI values at each time. A third means including an equation means for obtaining an interpolation polynomial for the predicted value, and an interpolation means for predicting a CSI value after an arbitrary time by using the interpolation polynomial, and outputting a predicted value as an interpolation value based on the prediction result. The rainfall / snowfall prediction device according to claim 2 or 3, further comprising: