JP3034167B2 - Weather disaster forecasting method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a meteorological disaster prediction method using a neural network, and more particularly to a method capable of highly accurate prediction even at the beginning of operation, and further improving the prediction accuracy with operation. It relates to a possible meteorological disaster prediction method .

[0002]

2. Description of the Related Art In today's highly electrified world, the supply of electricity must not be interrupted even every moment. If the supply of electricity is interrupted, it will cause enormous damage from general households to a wide range of industries. For this reason, when an accident (disaster) occurs on a transmission line, a maintenance management system that immediately detects the damage point and quickly recovers from the disaster has been developed and put to practical use.

[0003] In recent years, a predictive maintenance system has been developed which predicts the occurrence of a disaster by using weather data and the like, rather than knowing the disaster as a result, and prevents the disaster before it occurs. For example, in winter, snow damage caused by snow on transmission lines may cause widespread damage, and a snow damage prediction system for predicting the occurrence of such snow damage has been considered.

This snow damage prediction system uses a neural network to determine whether snow damage will occur three hours ahead based on weather data supplied online by the Meteorological Office or a private weather service business. I have. The literature includes "Development of a Snow Damage Warning System for Transmission Lines Using Neural Networks" (Iida et al .; 1994, The Institute of Electrical Engineers of Japan,
164).

As shown in FIG. 2, the above conventional snow damage prediction system inputs precipitation, temperature, and wind speed as weather data to a snow damage prediction neural network device and outputs a snow damage warning. The snow damage prediction neural network device is trained so that the presence or absence of snow damage can be determined from weather data based on weather data at the time of a past disaster. Weather data
Numerical data for a total of 6 hours, consisting of forecast data for the past 3 hours including the current time and forecast data for the past 3 hours, can be used to predict the presence or absence of snow damage three hours ahead.

[0006]

By the way, in learning of the neural network, weather data when a meteorological disaster occurred in the past is used as teacher data with a disaster. Further, although no meteorological disaster occurred, heavy snow fell. Weather data at the time, and weather data when no weather disaster occurred but the weather layout and weather data pattern were similar to the weather layout and weather data pattern at the time of the weather disaster, that is, at the time of non-meteorological disaster Weather data as teacher data without disaster. As a result, it is possible to successfully identify a difference in the pattern of weather data that is difficult to determine whether a human sees a disaster or not. In fact, 100% correct decisions can be obtained for past disasters. In addition, during operation of the system, each time a new disaster occurs, the weather data is re-learned as teacher data with the disaster, so that the discrimination ability of the neural network can be further improved.

[0007] However, it is unclear what answer the neural network will give except for what is taught, and when the system is actually operated, the weather data that is considered to have a low possibility of snow damage is obtained. On the other hand, it may be determined that there is snow damage. For this reason, the prediction accuracy is low. When a disaster is actually determined but no disaster actually occurs, this is referred to as “missing”. In some cases, the potential for a disaster was very high, but fortunately it did not result in a disaster. In some cases, the pattern of weather data was not similar.

Therefore, an object of the present invention is to solve the above-mentioned problems, and to provide a meteorological disaster prediction method capable of performing highly accurate prediction even in the early stage of operation and further improving the accuracy of prediction with operation. It is in.

[0009]

In order to achieve the above object, the present invention uses weather data such as precipitation, temperature and wind speed at the time of a meteorological disaster in the past as disaster-affected teacher data, and uses it for heavy snowfall. We let the neural network learn the weather data at the time of non-meteorological disaster where the weather layout and weather data have similar pattern at the time of weather disaster but there is no weather disaster as disaster-free teacher data, and then the neural network together to predict the occurrence of meteorological disasters enter the weather data, the weather disaster prediction method of re-learning neural network as a disaster supervised data the weather data every time the weather disaster occurs, the meteorological data at the time of weather disasters Similarity between the weather data pattern stored in the database and input and the weather data pattern stored in the database Calculated, there is no actual occurrence of meteorological disasters while determining neural net that there disaster, yet when the similarity is lower than a predetermined value, but to re-learn the neural network the meteorological data for disaster unsupervised data.

[0010] When the neural network determines that there is a weather disaster, or when the similarity is higher than a predetermined value, a warning of occurrence of a weather disaster may be issued.

[0011]

According to the above configuration, if the weather data at the time of the past weather disaster stored in the database has a pattern similar to the input weather data, the similarity is high, and if not, the similarity is low. It will be. When the degree of similarity is high, if the judgment of the weather disaster of the neural network is "having a disaster", it is appropriate, but if "no disaster", it is contradictory. When the degree of similarity is low, "no disaster" is appropriate, but "disaster" is contradictory. Thereafter, when the predicted time has elapsed, it can be confirmed whether or not a weather disaster has actually occurred. When the judgment of weather disaster is "has disaster",
If a meteorological disaster does occur, the results do not conflict, but if they do not, they do. When the judgment of the weather disaster is "no disaster", there is no contradiction unless a weather disaster actually occurs, but if it does occur, there is a contradiction.

When the judgment of the weather disaster of the neural network is inconsistent with the degree of similarity or inconsistent with the actual weather disaster occurrence result, the neural network can be re-learned using the weather forecast data as teacher data. it can. In the present invention, when the degree of similarity is low, that is, when the pattern is not similar to the weather data at the time of the past weather disaster, it is determined that “there is a disaster”, and no disaster actually occurs. Then, the neural network re-learns the weather forecast data as teacher data of "no disaster", and after the re-learning, determines the same weather data as "no disaster". Therefore, even if the weather data having an unknown pattern is erroneously determined for the first time, correct determination can be performed from the second time. Then, by continuing the operation, the effect of re-learning is accumulated, and the accuracy of prediction can be further improved.

Although the above method eliminates erroneous judgments during operation, erroneous judgments may occur at the beginning of operation. If there is no disaster, the judgment of "having a disaster" is still unsatisfactory. However, it is not preferable to make the judgment of "having no disaster" when a disaster is likely to occur. In addition to notifying the occurrence of a meteorological disaster when becomes, regardless of the judgment of the neural network,
Even when the similarity is higher than a predetermined value, it is preferable to issue a weather disaster occurrence warning. Thus, even if “no disaster” is erroneously determined, a warning is issued based on the large similarity.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below in detail with reference to the accompanying drawings.

As shown in FIG. 1, a snow damage prediction system according to the weather disaster prediction method of the present invention comprises a snow damage prediction neural network device 1 and an accident (disaster) database device 2.
, A similarity comparison device 3 and a logical sum device 4.

The snow damage prediction neural network device 1 uses weather data such as precipitation, temperature and wind speed at the time of a snow damage disaster in the past as teacher data with a disaster, and has a pattern similar to the weather layout and weather data at the time of heavy snow or a disaster. The weather data at the time of non-disaster was learned in advance as disaster-free teacher data,
When the weather data is input, the presence / absence of a snow disaster is predicted and determined, and “disaster is present” is notified to the OR device 4. Further, at the time of operation, re-learning is possible using the teacher data with disaster and the teacher data without disaster.

The disaster database device 2 stores weather data at the time of a past weather disaster, retrieves the stored weather data, and retrieves weather data having a pattern most similar to the input weather data. Can be. At the time of operation, it is possible to add weather data at the time of a new weather disaster online or offline.

The weather data stored in the disaster database device 2 is based on the snow damage prediction neural network device 1 described above.
May be the same as the past weather data used for learning. Further, these weather data are composed of data for 6 hours like the input weather data for comparison with the input weather data.

The similarity comparison device 3 is a device that compares patterns of weather data extracted using a disaster database with input weather data and represents the similarity. The similarity is a numerical value of the similarity between two weather data patterns. The similarity is high when the patterns are similar and low when the patterns are not similar. The similarity, in advance, the threshold value A for judging not similar / and, B is set. Similarity comparison device 3 is adapted to notify the logical sum unit 4 when the similarity exceeds the threshold value B. On the other hand, the threshold value A is used to determine the relearning of snow prediction neural network device 1.

In this snow damage prediction system, whenever a snow damage disaster occurs, the snow damage prediction neural network device 1 always relearns the weather data as disaster-affected teacher data. Further, when the judgment of the snow damage prediction neural network device 1 for a certain weather data is inconsistent with the level of similarity of the similarity comparison device 3 or inconsistent with the actual snow damage disaster occurrence result, the weather forecast data is used as teacher data. The snow damage prediction neural network device 1 is re-learned. In this embodiment, for a certain meteorological data, if the determination is "Disaster present", the similarity is not exceeding the threshold value A, and, if the actual disaster has not occurred, the weather data Is re-learned as disaster-free teacher data.

The logical sum device 4 determines whether the snow damage prediction neural network device 1 notifies the user of the judgment of “having a disaster”,
Be any of the case where the degree of similarity from the similarity comparator 3 exceeds the threshold value B is notified and outputs a snow warning by a logical OR.

Next, the operation of the embodiment will be described.

In the snow damage prediction system shown in FIG. 1, the snow damage prediction neural network device 1 has learned the past weather data such as precipitation, temperature, wind speed, etc. as teacher data and is ready for operation. At the time of operation, a total of 6 data including forecast data of the past 3 hours including the current time and forecast data of 3 hours ahead
The numerical data for the time is the neural network device for snow damage prediction 1
Is input to The snow damage prediction neural network device 1 determines the presence or absence of a snow damage disaster, and when the occurrence of a snow damage disaster is predicted, the “OR” is notified to the logical sum device 4.

At the same time, the same weather data is input to the similarity comparison device 3. The similarity comparison device 3 compares the pattern of the weather data extracted using the disaster database device 2 with the input weather data and indicates the similarity. Also,
Notify the logical sum unit 4 when the similarity exceeds the threshold value B.
The degree of similarity exceeds the threshold value A, B means that the meteorological data pattern weather data at the time of past snow disaster very similar is detected.

[0025] In the case snow determination of the prediction neural network device 1 of the "disaster there", the similarity does not exceed the threshold value A, and, in the case of actual disaster did not occur, the weather data The learning of the snow damage prediction neural network device 1 is performed again as disaster-free teacher data. After the re-learning, the same weather data is determined to be “no disaster”. Therefore, even if the weather data having an unknown pattern is erroneously determined for the first time, correct determination can be performed from the second time. Then, by continuing the operation, the effect of re-learning is accumulated, and the accuracy of prediction can be further improved.

The logical sum device 4 determines whether the snow damage prediction neural network device 1 notifies the user of the judgment of “having a disaster”,
Be any of the case where the degree of similarity from the similarity comparator 3 exceeds the threshold value B is notified to output a snow warning by a logical OR. As a result, the snow damage prediction neural network device 1 determines “no disaster”, and even if the determination is erroneous, an alarm is issued based on the high similarity determined by the similarity comparison device 3 and alarm omission is eliminated. For the same weather data, the second and subsequent times, the snow damage prediction neural network device 1 makes a determination of "having a disaster", so that a warning is reliably issued.

[0027] It should be noted that, in the case where the judgment of snow damage prediction neural network device 1 is "disaster there", the degree of similarity has exceeded the threshold value A, however, if the actual disaster did not occur, the disaster Although the potential was very high, it was thought that the disaster did not result in a disaster, so it was not relearned. In addition, despite the similarity exceeds the threshold value A, in the judgment of snow damage prediction neural network device 1 is "no disaster", when the actual disaster did not occur, snow damage prediction neural network device 1 Is important because it is determined on the basis of "something" different from the similarity determined by the similarity comparison device 3 and is not a target of re-learning.

The above embodiment is a system for predicting snow damage of a transmission line mainly in winter. However, the meteorological disaster prediction method of the present invention is used for various systems for predicting weather disaster and preventing damage in advance. it can.

[0029]

The present invention exhibits the following excellent effects.

(1) Even if learning of the neural network is not complete at the start of operation, the accuracy of prediction can be improved with operation, and development of a weather disaster prediction system becomes easy.

(2) Not only the judgment of the neural network, but also the alarm is performed by using the similarity obtained from the database, so that a highly accurate prediction can be made even at the initial stage of operation by compensating for the disadvantage of the neural network with respect to unlearned data Thus, the omission of the alarm can be eliminated.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a snow damage prediction system according to an embodiment of the present invention.

FIG. 2 is a configuration diagram of a snow damage prediction system showing a conventional example.

[Explanation of symbols]

1 snow damage prediction neural network device (neural network) 2 accident (disaster) database device (database) 3 similarity comparison device 4 logical sum device

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-6-105428 (JP, A) Patent 2875944 (JP, B2) Patent 2726589 (JP, B2) (58) Fields investigated (Int. Cl. ⁷ , DB name) G01W 1/00-1/18 G05B 13/00-13/04 G06F 15/18 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. Weather data such as precipitation, temperature, wind speed, etc. at the time of a meteorological disaster that occurred in the past is used as disaster-affected teacher data. The weather data at the time of the non-meteorological disaster in which the weather disaster did not occur while having the neural network is learned as disaster-free teacher data, and then the weather data is input to the neural network to predict the occurrence of the weather disaster, A meteorological disaster forecasting method that retrains the neural data as a teacher data with disaster every time a meteorological disaster occurs. The degree of similarity to the pattern of the weather data obtained is determined, and the neural network determines that there is a disaster while executing Of no occurrence of weather disasters, yet when the similarity is lower than a predetermined value, weather disasters prediction method characterized by re learning neural network the meteorological data for disaster unsupervised data. 2. The meteorological disaster according to claim 1, wherein when the neural network determines that a meteorological disaster has occurred or when the similarity is higher than a predetermined value, a warning of occurrence of a meteorological disaster is issued. Forecasting method .