JPH0458188A - Weather forecasting device - Google Patents

Abstract

PURPOSE:To accurately forecast weather by using information regarding the actual weather as an expected value and varying the weight of a neural network so that the expected value is outputted when weather condition data at specific time are inputted to the neural network. CONSTITUTION:The neural network 3 receives data inputted from memories 21 - 24 at an input layer 31 and outputs them to an output layer 33 through an intermediate layer 32. Specific weight is given among those layers 31 - 33, and properly varied and adjusted to a specific value by a weight variation quantity computing element 34. For this weight variation and adjustment, the information regarding the actual weather is supplied from a rainfall sensor 15 to the weight variation quantity computing element 34. The information regarding the weather forecast by the neural network 3 is outputted from the output layer 33 to a display 4 and displayed.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a weather prediction device capable of predicting the weather.

[Conventional technology]

The weather forecasting device is placed, for example, inside a station.

This conventional weather forecasting device detects changes in atmospheric pressure, predicts the weather after a certain period of time, and displays the forecast. usually,
When the atmospheric pressure is rising, sunny skies are predicted, and when it is falling, rain is predicted.

[Problem to be solved by the invention]

However, since conventional devices predict the weather only based on atmospheric pressure, they have been unable to improve the accuracy of their predictions beyond a certain level.

The present invention has been made in view of this situation, and aims to make it possible to predict the weather more reliably.

[Means to solve the problem]

The weather prediction device of the present invention includes a processing means comprising a neural network having an input layer, an intermediate layer, an output layer, and a weight change amount calculator, and at least one of weather conditions such as atmospheric pressure, temperature, and humidity. A detection means for detecting, a storage means for storing weather conditions detected by the detection means and outputting them to an input layer of the processing means, and a supply means for supplying information regarding the actual weather to a weight change amount calculator of the processing means. ,
The apparatus is characterized by comprising display means for displaying weather forecast information output from the output layer of the processing means.

[Effect]

In the weather forecasting device configured as described above, information regarding actual weather is input to the neural network as an expected value. Then, when a predetermined data among the data stored in the storage means is input, the weight is changed by the weight change amount calculator so that the expected value thereof is output. Therefore, this learning enables accurate weather prediction.

〔Example〕

FIG. 1 is a block diagram showing the configuration of an embodiment of the weather forecasting device of the present invention.

Reference numeral 1 denotes a sensor as a detection means for detecting various weather conditions. In this embodiment, atmospheric pressure sensor 11, temperature sensor 12, humidity sensor 13, and other sensors 14 detect atmospheric pressure, temperature, humidity, and other conditions. weather conditions are now detected. These sensors 11
The outputs of 14 are supplied to the memories 21 to 24 (storage means) of the prediction unit 2, respectively, and stored therein. memory 21
The data stored in 24 are input to the neural network 3.

The neural network 3 receives data input from the memories 21 to 24 at an input layer 31, and outputs the data to an output layer 33 via an intermediate layer 32. A predetermined weight is assigned between each of these layers, and this weight is adjusted to a predetermined value as appropriate by the weight change amount calculator 34. To adjust this weight change, the rain sensor 15 (
Information regarding the actual weather is supplied to the weight change amount calculator 34 from the supply means (supplying means).

Information regarding the weather predicted by the neural network 3 is output from the output layer 33 to the display 4,
Is displayed. In this embodiment, the predicted weather is displayed using one of the graphics 41 to 43 representing sunny, cloudy, or rainy weather, respectively.

This display can also be displayed as a probability of rain using the number 51, as shown in FIG. 2, for example.

Next, the neural network 3 will be further explained.

As shown in FIG. 3, the neural network 3 is composed of L layers from the first layer to the second layer.

The first layer is an input layer 31, the second layer is an output layer 33, and the layers from the second layer to the (L-1)th layer are intermediate layers 32.

N(m) (m=1.2.3...L) neurons (indicated by @ middle circle) are arranged in the m-th layer, and the (m-1)
The first neuron (i = 1.2.3...N (1)) in the layer and the jth neuron in the mth layer have weights Wij (m-1) (the electrical resistance diagram in the figure). ).

The symbols in the figures have the following meanings.

xi: i-th input dj: j-th expected output (j=1.2.3...N (
L) ) yj (m): Output of the j-th neuron of the m-th layer (j
=1.2.3...N(m) The output function of the first layer is yi (
1)=fl (xi)=xi i=1.2,...,N(1) It is a linear function expressed by =1.2, . . ., N(m) m=2, 3, . . ., L It is a sigmoid function. Here, netj
(m) L; t,, Input of the jth neuron of the mth layer, i = 1.2 according to the definition of the neural network,
・ ・ n, N (m-1)j=1.2, ・ ・ ・
, N(m) m=2, 3, . . . , L.

The input xi and the expected output dj that should be obtained from that input are
When presented to the neural network, the neural network processes the input xi with the current connection weights between neurons and transforms it into an output yj (L). Next, the weight change amount calculator 34 compares this output yj(L) with the expected output dj, finds the difference between the two, and changes the weight of the connection between neurons so that this difference becomes smaller. . 34, in pack propagation, j j=1. 2. The weights of connections between neurons are changed so that the evaluation function E represented by N(L) becomes smaller.

Here, a wij(m-1) a netj(m)
a wij(m-1)=-dj(m) ・yi(m-
1) i=1.2, ... , N(m-1)j=1.2,
・ ・ ・, N(m) m, = 2, 3, ・ ・ , When defined as L, the i-th neuron of the (m-1)th layer and the j of the m-th layer
The correction amount Δwij (m-1) of the connection weight wij (m 1) of the th neuron is expressed as Δwij (m-1)=c・dj (m)・yi(m-1
), then the weight of the current connection between neurons wij(
The evaluation function E can be minimized by adding the correction amount Δwij(m-1) to m-1) and correcting it to a new weight of connection between neurons wij(m-1).

wij(m-1)(t+1)”wij(m 1)(t
)+Δwij(m 1)(t) 1=0, 1, . . . where C is a learning parameter, dj (m) is an error, and t is a discrete time. In this case, t increases by 1 each time learning is completed.

In the case of the second layer, the error δj (L) is a E a E a yi (
L) = (dj-yi(L) ・f2' (net
j(L))j=1.2, . . ., N (L). In addition, in cases other than the th layer, the error δj (m) is Ka netk (+++1) a yi
(m)・f 2' (netj(+++)) m=2, 3, . . . , L.

In the case of neural networks, self-learning is possible. Next, this procedure will be explained according to the pack propagation learning algorithm.

(1) Input xi to the input layer.

(2) Calculate the output for the input xi sequentially from the input layer 31 to the output layer 33 to obtain the final output yj(L).

yi(1)=f 1(xi) i=1.2,..., N(1) j=1.2.・ ・ ・, N (m) k=1, 2
．．・ ・ +, N (m+1)m=2, 3.・
・ ・ , becomes L-1. Here, since the output function is a sigmoid function, f2' (natj(m)) = f2(netj(m))
・(1-f2(netj(m)))j=1.2, ・
・ , N (m)yj (m)=f2 (ne
tj (m))i = 1, 2. --- , N(m-
1) j=1.2, ・ ・ , N (m)m =
2, 3. . . . , N(3) Present the expected output dj to the output layer 33.

(4) From the expected output di and output yj (L), calculate the error δj (L) of the second layer, and calculate the weight wi of the connection to the output layer 33.
Modify j(L-1).

δj (L) = (dj−yj (L))・f 2
'(netj (L))Δwij(L-1)=c
・δj(L) ・yi(L-1)wij(L-1)
(t+1)=vij(L-1)(t) Δwij(L-
1)(t)i=1.2. ..., N (L-1) j=1.2, ..., N (L) 1=0.1, ..., where C is a learning parameter and t is a discrete time.

(5) From the error δk(m+1) of the (n+1)th layer, the mth
Find the error δj(m) of the layer, and calculate the weight wi of the connection to the mth layer.
Modify j (m-1).

Δvij (m-1) = c・δj (m)・yi (votes-1) wij (m-1) (t 41) = wij (m
-1) (t) Δwij (m-1) (t) i=1.2
,...,N(m-1) j=1.2, . . .,N(m) k=1.2,...,N(m+1)m==2
, 3, ・ ・ , L−1 However, the error δj of the first layer
When calculating (m), the weight wjk(m
).

(6) Return to the process in (1) above and repeat the correction of the weights between neurons until the evaluation function E reaches an appropriate value.

When the evaluation function E reaches an appropriate value, the patterns of the input xi and expected output dj are changed and these steps are repeated.

Further, the operation of this learning mechanism will be explained in detail with reference to FIG. However, as shown in Figure 4, the output function of the neurons in the first layer is a linear function expressed as yi (1) = f 1 (xi) = xi i = 1.2, ..., N (1) The output functions other than the first layer are: j=1.2, . . ., N(m) 1=2, 3. . . . is a sigmoid function represented by L.

(1) When inputs x1 and x2 are input to neurons A and B, outputs yl (1) and y2 (1) are output from neurons A and B via the output function.

y 1 (1) = f I Cx 1) y2 (1) =
f1 (χ1) (2) Neuron C has the output power y of neurons A and B.
1, y2 are the respective connection weights Wll, W21 (1
), and their sum entl(2) is input. Similarly, neuron D has a connection weight W12 (1) for the outputs y1 and y2 of neurons A and B, respectively.
, W22 (1), the sum of the values net2 (2) is input.

netl(2)”=yl(1)・Wll(1)+y2(
1)・W21(1) net 2(2)=y 1(1)・W12(1)+y
2(1)・W22(1) Neurons C and D have inputs netl(2) and net2(
2) is processed with an output function to produce an output y 1 (2).

Output y2(2).

y 1(2) = f 2 (net 1(2))
) y 2 (2) = f 2 (net 1 (2)
) ) (3) For neuron E, the outputs yl(2) and y2(2) of neurons C and D have respective connection weights W.
The sum of the values multiplied by ll(2) and W21(2) netl
(3) is input.

net 1(3)=y 1(2)・Wll(2)+
y2(2)/W21(2) Neuron E processes input net 1(3) with an output function and outputs output yl(3).

y 1 (3) = f 2 (net 1 (3)) (4)
The weight change amount calculator DI calculates the error δ1(3) of the neuron E from the output y1(3) of the third layer and the expected output d1.

δ1(3)=(d 1 y 1(3)) ・f 2
(net 1(3)) ” (1-f 2(ne
t 1 (3)) its error δ1 (3) and neuron C
From the output y1(2), neuron C, El'1lfi
A correction amount ΔWll(2) of the connection weight is calculated.

ΔWll (2)=C·δ1 (3)·yl(2) However, C is a learning parameter. Then, the weight correction amount ΔWll(3) is added to the weight Wll(2) of the connection, and the weight Wll(2) is corrected to a new weight.

Wl 1(2)(t+1)=W11(2)(t)+ΔW
l 1(2)(t) t=o, l ・ ・ ・ However, t is a discrete time.

(5) Similarly, the weight change amount calculation unit D2 also uses the neuron D.
, the correction amount ΔW21 (2) of the weight of the connection between E.

ΔW21 (2) = C・δ1 (3)・y2(2) Then, add the correction amount ΔW21 (2) of the weight of that connection to the weight W21 (2) of that connection, and change the weight 21 (2) to the new weight. Modify the weight.

ΔW21(2)(t+1)=W21(2)(t)+ΔW
21(2)(t) (6) The weight change amount calculator D3 is the weight change amount calculator D1
The error δ1(2) of the neuron C is calculated from the error δ1(3) of the neuron E calculated in step 1 and the weight Wll(2) of the connection between the neurons C and E.

δ1(2)=61 (3) ・Wl 1 (2) ・f
2(n etl(2))” (1-f 2(n e
t 1 (2)) However, at this time, the weight W before correction
Use ll(2). From the error δ1(2) and the outputs yl(1) and y2(1) of the neurons A and B, the correction amount ΔW21(1) of the weight of the connection between the neurons A and C,
The amount of correction ΔW21 (1
) is calculated.

ΔWll(1)=C・δ1(2)・yl(1)ΔW21
(1)=C・δ1(2)・y2(1) Then, change the weight correction amounts ΔW11(1) and ΔW21(1) of those connections to the weights Wll(1) and W21(1) of those connections. and modify the weights Wl 1 (1) and W21 (1) to new weights.

wi 1 (1) (t+1) = wl 1 (1) (t) ten Δ
Wl 1(1)(t) W21(1)(t+1)=W2B1)(t)+ΔW21
(1) (t) (7) The weight change amount calculator D4 is the weight change amount calculator D2.
The error δ2(2) of the neuron D is calculated from the error δ1(3) of the neuron E calculated in the above and the weight W21(2) of the connection between the neurons D and E.

δ2 (2)=61(3)・W21(2)・f2(ne
t 2(2)) ・(1-f 2(net 2(2)
)) However, at this time, the weight W21(2) before modification is used. From the error δ2(2) and the outputs yl(1) and y2(1) of neurons A and B, the correction amount ΔW12(1) of the connection weight between neurons A and 0 and neuron B
, D is calculated.

ΔW12(1)=C・δ2(2)・yl(1)ΔW22
(1)=C・δ2(2)・y2(1) And the amount of modification of the weight of those connections ΔW12(1), ΔW22(1)
is added to the weights Wl2(1) and W22(1) of those connections, and the weights Wl2(1) and W22(1) are modified to new weights.

Wl 2B)(t+1)=W12(1)(t)+ΔW1
2(1)(t) W22(1)D+1)=W22(1)(t)+ΔW22
(1) (t) By repeating these learning processes, the output yl
(3) is corrected to the expected output d1.

In this way, by self-learning, the neural network can acquire the function of a pattern conversion device that converts an input pattern into an output pattern.

The memories 21 to 24 in FIG. 1 store data detected by the sensors 11 to 14 a predetermined time ago. Therefore, the actual weather information outputted by the precipitation sensor 15 is input to the weight change amount calculator 34 as the above-mentioned expected value, and at the same time, from the time when this actual weather information occurs (when it rains), 1, for example, 1 When data from the previous hour is read from the memories 21 to 24 and input to the weight change amount calculator 34 for self-learning, the weights are corrected to be suitable for predicting the weather one hour later. This makes it possible to realize a prediction device that predicts the weather one hour later.

Of course, this time can be set to any value such as 2 hours, 8 hours, 24 hours, or one week.

〔Effect of the invention〕

As described above, according to the weather prediction device of the present invention, when information regarding actual weather is used as an expected value and weather condition data at a predetermined time stored in the storage means is input to the neural network.

In order to output the expected value, we changed the weight of the neural network using the weight change amount calculator, so this learning makes it possible to accurately predict the weather after an arbitrary amount of time. .

[Brief explanation of the drawing]

Figure 1 @ is a block diagram showing the configuration of an embodiment of the weather forecasting device of the present invention, Figure 2 is an explanatory diagram of the display display, Figure 3 is a diagram showing the configuration of the neural network, and Figures 4A and B. 5 is a characteristic diagram of the output functions of the input layer and layers other than the input layer in the neural network, and FIG. 5 is a diagram explaining the self-learning function of the neural network. 1...Sensor (detection means), 3...Neural network (processing means), 4.Display (display means)
, 15... Rainfall sensor (supply means), 21 to 24.
...Memory (storage means), 31...Input layer, 32...
- Middle layer, 33... Output layer, 34... Weight change amount calculator. Patent applicant OMRON Corporation

Claims (1)

[Claims] Processing means comprising a neural network having an input layer, an intermediate layer, an output layer, and a weight change amount calculator, and a detection device for detecting at least one of weather conditions such as atmospheric pressure, temperature, and humidity. means for storing the weather conditions detected by the detecting means and outputting them to an input layer of the processing means; and supply means for supplying information regarding the actual weather to the weight change amount calculator of the processing means. A weather prediction device comprising: and display means for displaying weather prediction information output from the output layer of the processing means.