JP3250601B2 - Weather forecasting device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for calculating a motion vector from a weather image and predicting a coefficient to be set for the motion vector by a neural network model or the like.
The present invention relates to a weather prediction device that predicts a weather image at a future time.

[0002]

2. Description of the Related Art A weather image includes a weather satellite image, a radar AMeDAS composite diagram, and the like. In the following description, use of a weather radar image will be described as an example. Conventionally, a cross-correlation technique has been used as a technique for predicting a future weather radar image from a past weather radar image. In this method, first, a cross-correlation value is obtained from the following two equations from two radar images separated by a certain time interval Δt. Here, the two radar images are R1 and R2, and the gradation of the image at the grid point (i, j) on the radar image, that is, the rainfall snowfall intensity is (R1, i, j), (R2, i, j), the area where the correlation is obtained is (M, N), and the deviation between the two radar images when calculating the correlation value is (k, l) (FIG. 4).
The oblique line indicates the range in which the correlation value is obtained, and the thick arrow at the center indicates the moving direction of the rain cloud.) (Yuo Yuma, Katsuhiro Kikuchi,
Imahisa: "On the echo moving speed by a simple weather radar", Hokkaido University Geophysical Research Report, Vol. 4
4, pp. 23-34, 1984. )

[0003]

(Equation 1)

[0004]

(Equation 2)

[0005]

(Equation 3) The cross-correlation value obtained by the calculation is as shown in FIG. Therefore, interpolation using a quadratic function is performed on the point (K, L) on the grid point at which the maximum value of the cross-correlation value is obtained and the cross-correlation values of the four points in the vicinity thereof, and the cross-correlation value is determined The deviation (k ', l') is calculated by the following equation (FIG. 6; however, only the X component is shown). However, the maximum correlation value on the grid point and the cross-correlation values of the four neighboring points are σ _KL , σ _-x , σ _{+ x} , σ _-y , and σ _{+ y} , respectively.

[0006]

(Equation 4)

[0007]

(Equation 5) As described above, when the two images are shifted by (K + k ′, L + 1 ′), the cross-correlation value becomes maximum. From this,
The moving vector of the rain cloud is obtained by the following equation.

[0008]

(Equation 6)

[0009]

(Equation 7) Here, assuming that the interval between the two images for which the movement vector was obtained is Δt and the time to be predicted is t, the obtained movement vector of the rain cloud is multiplied by a coefficient of (t ÷ Δt) to obtain 30 minutes to 3 hours. By calculating the amount of movement and moving the radar image in parallel by the amount of movement, a radar image 30 to 3 hours ahead was predicted (FIG. 7).

[0010]

However, the moving speed in the rainfall and snowfall area is not constant, that is, the coefficient is rarely constant (t ÷ Δt), and the coefficient varies with time in most weather conditions. In addition, since it differs depending on the time interval to be predicted, a simple extrapolation method that only determines the coefficient in proportion to the prediction time, as in the conventional cross-correlation method, reduces the prediction accuracy one hour or more ahead. There is a problem. FIG. 3 shows an example in which an image three hours ahead of a certain radar image is predicted in order to examine how the prediction accuracy changes when the coefficient is continuously changed. The vertical axis indicates the sum of squared errors between the predicted image and the real image, the horizontal axis indicates the coefficient value, and the vertical line at the center indicates the coefficient value (tｔΔt) used in the conventional cross-correlation method. From this figure, it can be seen that there is a minimum value of the error, that is, an optimum value of the coefficient, at a place smaller than the coefficient set by the cross-correlation technique. Therefore, an object of the present invention is to provide a weather forecasting device based on a weather forecasting method in which a coefficient to be set in a movement vector is predicted by a neural network model or the like and the prediction accuracy at least one hour ahead is improved.

[0011]

According to the present invention, there is provided a weather forecasting apparatus comprising: an input unit including a weather radar for measuring a rainfall and snowfall area; and a file reading unit for reading parameters required for weather forecasting. A movement vector calculation unit that calculates a movement vector of a rainfall snowfall area based on information from the input unit, and a time system of a coefficient by which the movement vector is multiplied
A coefficient calculator for calculating a sequence change, the coefficient prediction unit for predicting learning the time series change of the time-series change one time destination of said coefficients as a teaching value, the most recent rate transferred from the input unit
The motion vector calculated by the motion vector
In the direction of the vector, and further by the coefficient prediction unit
Radar image of future time with image multiplied by predicted coefficient
A prediction image generation unit for generating a radar image, and an output unit for displaying the radar image at the future time.

Further, the movement vector calculation unit provides a weather forecasting apparatus for calculating a movement vector of a rainfall snowfall area using a cross-correlation technique.

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Weather radar 10 of input unit 100
1 is transferred to the processing unit 200. Also, values such as parameters necessary for learning and prediction are transferred from the file reading device 102 to the movement vector calculation unit 2.
Transfer to 01. The movement vector calculation unit calculates the movement vector of the rainfall snowfall area by the cross-correlation technique, and sends it to the coefficient calculation unit 202. The coefficient calculation unit calculates a coefficient by which the movement vector is multiplied and transfers the coefficient to the coefficient prediction unit 203. In the coefficient prediction unit, an optimal coefficient at the time of prediction of a radar image or the like is obtained using a neural network model, and this coefficient is transferred to the prediction image generation unit 204. The prediction image generation unit multiplies the latest radar image from the input unit 100 by the coefficient transferred from the coefficient prediction unit 203 in the direction of the movement vector transferred from the movement vector calculation unit 201, and calculates the predicted radar image after a predetermined time. The image is transferred to the output unit 300 as an image.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the weather forecasting apparatus described in the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the apparatus when a weather radar image is used, in which 100 denotes an input unit, 200 denotes a processing unit, and 300 denotes an output unit. The input unit 100 includes a weather radar 101 for measuring a rainfall / snowfall region, a file reading device 102 for reading information necessary for calculating a movement vector by a cross-correlation technique, calculating a coefficient, and predicting a coefficient.
The processing unit 200 includes a movement vector calculation unit 201 that calculates a movement vector of the rainfall and snowfall region by a cross-correlation technique, a coefficient calculation unit 202 that calculates a coefficient by which the movement vector is multiplied,
A coefficient prediction unit 203 for predicting a time-series change of a coefficient, and a predicted image generation unit 204 for creating a radar image at a future time using a movement vector at an arbitrary time and a predicted coefficient. In this embodiment, the coefficient prediction unit 203 predicts a coefficient using a neural network model.

A radar image measured using the weather radar 101 of the input unit 100 is input and transferred to the processing unit 200. Also, values such as parameters necessary for prediction of a weather radar image are transferred from the file reading device 102 to the movement vector calculation unit 201, the coefficient prediction unit 203, and the predicted image generation unit 204.

The movement vector calculation unit 201 includes the input unit 100
Of the two radar images at the time interval Δt transferred from the
x, Vy only staggered, time t + Vx when the cross-correlation value is maximized with the radar image of Delta] t, seeking Vy, the movement vector of rainfall snowfall area at time t V _t
= (V _x , V _y ), and similarly at time t + Δt, t
+ 2Δt,...,

[0017]

(Equation 8) Are calculated as many as the number is set arbitrarily in time series, and are transferred to the coefficient calculation unit 202.

The coefficient calculating unit 202 converts the radar image at time t transferred from the moving vector calculating unit 201 into a moving vector V _t = (V _x , V _y ) at that time also transferred from the moving vector calculating unit 201. Is shifted by multiplying it by a continuously changing coefficient α _{l, t} in the direction of, and the sum of squared errors between the image and the image one hour later is obtained. The coefficient α _{l, t} when the error sum of squares is minimized is used as a coefficient for prediction one hour ahead. This calculation is performed at time t + Δt, t + 2Δ
The same applies to t,.

[0019]

(Equation 9) That is, the time series change of the coefficient is calculated and transferred to the coefficient prediction unit 203. Here, 1 indicates a prediction time interval. Coefficient prediction unit 20
No. 3 learns the time-series change of the coefficient α transferred from the coefficient calculation unit 202 using α one time ahead as a teacher value.

Here, an example of a neural network model used in the present invention will be described. Here, a hierarchical neural network model is used as a typical example, but other types of models such as a neural network model having a regression connection and a polynomial approximation can also be applied. FIG. 2 is a diagram illustrating an example of a hierarchical neural network model. The hierarchical neural network model is a layered network model including one input layer, a plurality of intermediate layers, and one output layer, and each layer is composed of an appropriate number of units.
Each layer unit except the input layer receives the weighted input from the unit of the previous layer, calculates the sum thereof, and outputs a result obtained by multiplying the sum by an appropriate function f. That is, X ^k _i , O ^k _i
(I = 1, 2,..., L, L: the number of units in the previous layer) are respectively the sum of inputs and the output of the i-th unit in the k-th layer,
Assuming that W ^k−1 _{j, i} is the coupling strength from the j-th unit in the k−1-th layer to the i-unit in the k-th layer, the following equation is obtained.

[0021]

(Equation 10) A typical sigmoid function (Equation (9)) is used as the input / output function f of the unit, but another conversion function may be used depending on the model.

[0022]

[Equation 11] The learning algorithm at this time is based on the error backpropagation method (Japan Industrial Technology Promotion Association, Neurocomputer Research Group: “Basics of Neurocomputing”, pp. 4).
8-50).

When trying to make predictions one hour ahead, the movement vector of the radar image at the start of the prediction is obtained from the movement vector calculation unit 201, but the coefficient α transferred from the coefficient calculation unit 202 is the latest one. But you can only get one hour ago. Therefore, an arbitrarily specified number of coefficients α transferred from the coefficient calculation unit 202 to the neural network model of the coefficient prediction unit 203 is input, and the arbitrarily specified number of coefficients α at a time earlier than that is used as teacher data. Do the learning. The latest arbitrarily specified number of coefficients α is input to the trained neural network model, the coefficient α at the previous time is output, the coefficient α is input again to the neural network model, and this is repeated. Thus, the coefficient at the start of the prediction is calculated. This repetition processing will be described with reference to FIG. The output O ^k _i of the neural network model is an arbitrary number of input coefficients α (FIG. 2).
Is a predicted value at the time before X _{1 to} X _L ), the predicted value is again brought to the input side, and the coefficient α is sequentially updated by passing through the delay operator D and input to the input layer. By repeating this, the coefficient α at the next time can be obtained sequentially. The calculated coefficient α at the start of the prediction is transferred to the predicted image generation unit 204.

The predicted image generation unit 204 shifts the latest radar image transferred from the input unit 100 by multiplying by the coefficient α transferred from the coefficient prediction unit in the direction of the movement vector transferred from the movement vector calculation unit 201. The resulting image is transferred to the output unit 300 as a predicted radar image one hour later. The output unit 300 displays the predicted radar image transferred from the predicted image generation unit 204 on a display or the like.

[0025]

By using the present invention in prediction of weather images such as weather radar images, the present invention can be compared with a conventional simple extrapolation method in which a coefficient is determined in proportion to the prediction time.
More accurate weather image prediction becomes possible.

[Brief description of the drawings]

FIG. 1 is a block diagram of a weather prediction device according to the present invention.

FIG. 2 is a diagram showing an example of a hierarchical neural network model.

FIG. 3 is a diagram showing a change in a prediction error accompanying a change in a coefficient.

FIG. 4 is a diagram showing a cross-correlation technique.

FIG. 5 is a diagram showing an example of a correlation value distribution obtained by a cross-correlation technique.

FIG. 6 is a diagram illustrating a method of calculating a point k ′ at which a maximum value is obtained by quadratic interpolation from a correlation value distribution.

FIG. 7 is a diagram showing a prediction method using a cross-correlation technique.

[Explanation of symbols]

 REFERENCE SIGNS LIST 100 input unit 101 weather radar 102 file reading device 200 processing unit 201 movement vector calculation unit 202 coefficient calculation unit 203 coefficient prediction unit 204 predicted image generation unit 300 output unit

──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Noboru Sonehara 1-6-1, Uchisaiwaicho, Chiyoda-ku, Tokyo Nippon Telegraph and Telephone Corporation (72) Kazuo Abe 1-1-6 Uchisaiwaicho, Chiyoda-ku, Tokyo Japan Telegraph and Telephone Co., Ltd. (72) Inventor Masaharu Fujii 3-3-1 Kita-Jo Nishi, Chuo-ku, Sapporo, Hokkaido Examiner, Sapporo General Information Center Co., Ltd. Toru Hongo (56) References JP-A-7-120563 (JP) , A) JP-A-6-347563 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G01W 1/10 G06F 19/00

Claims (2)

(57) [Claims] 1. An input unit (100) comprising a weather radar (101) for measuring rainfall and snowfall area, and a file reading unit (102) for reading parameters required for weather prediction. A movement vector calculation unit (20) that calculates a movement vector of a rainfall snowfall area based on information from (100);
1), a coefficient calculation unit (202) for calculating a time series change of a coefficient multiplied by the movement vector, and a coefficient prediction for learning the time series change of the coefficient by learning the time series change one time ahead as a teacher value. Unit (203) and the input unit (100)
The latest radar image transferred from
Moved in the direction of the movement vector obtained by (201)
And the coefficient is predicted by the coefficient prediction unit (203).
Predicted image generating unit for raw <br/> formed an image obtained by multiplying the coefficients as a radar image of a future time (204) from the configured processor (200), an output unit for displaying the radar image of the future time (30
0). 2. The weather prediction device according to claim 1, wherein the movement vector calculation unit (201) calculates a movement vector of the rainfall snowfall area using a cross-correlation technique.