KR101610595B1 - Expedient Storm Surges Prediction System and Method - Google Patents

Abstract

The present invention relates to a system for predicting emergent storm surges, and to a method. The system generates a predicted model with a multi-layer structure by using a weather variable and a sea weather variable of a storm, and improves the accuracy of the predicted model through repeated learning, by performing the repeated learning with respect to the predicted model and correcting the predicted model. In addition, the system is capable of rapidly predicting the same with respect to weather data by using the predicted model without a complicated calculation process, and predicting the storm surges in a rapid and accurate manner with respect to a changing environment, by applying the latest data and updating the predicted model, thereby having effects in preventing the damage due to storm surges in advance.

Description

{Expedient Storm Surges Prediction System and Method}

The present invention relates to a rapid storm tsunami prediction system and method, and more particularly, to a storm tsunami prediction system and method for predicting coastal tsunami in order to minimize storm surge damage.

Due to recent global warming and weather changes, storm surge damage caused by typhoons is increasing. In particular, coastal cities are exposed to tsunami-related tsunami hazards and the damage is also very large. The damage caused by the typhoon is large, but it is difficult to predict the damage range, the target area, and the flooding depth when a tsunami occurs.

Generally, tidal currents generated by typhoons have been calculated mainly in areas where damage has been done through post evaluation, and the weather conditions such as the path of the typhoon, the central pressure, and the maximum wind radius vary greatly.

In addition, a system for predicting tsunamis during storm surges and a database of tsunami caused by typhoons that have occurred in the past are not sufficient, so they are not able to exert their original functions for urgent tsunami countermeasures.

Numerical experiments have been conducted to predict the intensity and frequency of typhoons and the intensity of storm surges. Numerical models have been constructed using existing data and many studies have been conducted to predict the tidal currents based on them. However, existing numerical models are not able to satisfy both accuracy and speed simultaneously.

Especially, it is possible to apply the developed numerical model to the rapid storm surge prediction, but it is not suitable for the Korean situation or the accuracy of the prediction result is low.

On the other hand, when the accuracy is improved, there is a problem that the calculation cost is large and the process of processing the data takes a lot of time.

In addition, although the accuracy of the result can be improved by newly reflecting the new data to the numerical model and building a new prediction model, it takes a long time to apply the new data, and thus it can not cope with the changing situation.

It is an object of the present invention to provide a storm tsunami prediction system and a storm tsunami prediction system that can predict a storm tsunami quickly and accurately and apply a latest data to update a prediction model to improve the accuracy of a prediction result, And to provide a tsunami prediction system and method therefor.

The storm tsunami prediction system according to the present invention includes a database (DB) for storing weather variables, sea parameters, and observation data of a typhoon, an initial prediction model using weather variables and sea parameters stored in the database, A predicted model generating unit for generating a predictive model by repeatedly learning the initial predictive model, a prediction result generation unit for generating a predictive result for input weather data using the predictive model generated from the predictive model generation unit, Wherein the prediction model generating unit includes a model building unit that generates the initial prediction model in a feedforward manner of a multi-layer structure including an input layer, a hidden layer, and an output layer, The weight of the initial prediction model of the initial prediction model and the weight By updating the value by modifying the initial prediction model includes a learning for generating the prediction model.

In addition, the prediction model generation unit may construct the neural network and generate the prediction model through iterative learning on the initial prediction model while changing the number of neurons.

In addition, when new data is generated, the prediction model generation unit may update the previously generated prediction model by applying the new data, or may generate a new prediction model.

In addition, the weather variables of the typhoon are latitude, longitude, center pressure, direction of progress, speed of progress, maximum wind radius, and the sea level variable of the typhoon is water level, waves, . ≪ / RTI >

Also, a communication module; And a control unit for controlling input / output of data, and the control unit controls the output of the prediction result through the output unit and transmits the result to an external terminal or a disaster warning system through the communication module.

The model building unit may set a number of neurons so that the number of neurons is two to three times the number of input variables of the input layer for the hidden layer, and weighting (W) and convenience (W) for the hidden layer and the output layer, bias, b) may be applied to generate the initial prediction model.

The learning unit divides the data stored in the database into learning data, verification data, and test data in an arbitrary division manner according to a predetermined ratio, and uses the data for learning and verification. The learning data, the verification data, So that the independence can be maintained without overlapping.

In addition, the learning unit terminates the learning when the result value satisfies the target value during the learning, and changes the number of the neural numbers when the result does not satisfy the correlation coefficient, repeats the data division and repeats the learning have.

The learning unit may repeat the learning of the predictive model based on the learning data. When the result derived from the learning data satisfies the target value, the learning unit may use the verification data or the test data for the learning data, It is possible to verify the prediction model learned by the learning data.

According to another aspect of the present invention, a prediction method of a storm surge prediction system according to the present invention includes the steps of generating an initial prediction model of a multi-layer structure using a weather parameter and a sea parameter of a typhoon, A step of modifying the weight and convenience of each layer of the initial prediction model through an iterative learning on the model, and a step of, if the result of the prediction model corrected by the learning satisfies the target value, And setting the prediction model as a final prediction model; and when the weather data is input, applying the prediction model to output a prediction result.

The storm tsunami prediction system and method according to the present invention as described above can improve the reliability of the prediction model through learning while generating a prediction model through modeling from existing data and update the prediction model by applying the latest data Rapid and accurate predictions of storm surges can be made in accordance with the changing environment, thus preventing damage from storm surges in advance.

In addition, it is possible to output a quick prediction result in response to a change in the course of a typhoon or the like, which is converted in real time, and to generate a time-series prediction result as well as a regional prediction result.

In addition, it is easy for non-professional users to check the prediction result, so it is possible to make a quick decision based on the prediction result in a rapidly changing situation, so that the efficiency is greatly improved.

1 is a schematic diagram of a prediction method of a storm surge prediction system according to the present invention.
2 is a diagram illustrating a configuration of a storm surge prediction system according to the present invention.
3 is a diagram referred to explain a prediction model generation of the storm surge prediction system of the present invention.
4 is a graph showing learning and verification in predictive model learning of the present invention.
5 is a flowchart illustrating an operation method of a storm surge prediction system according to the present invention.
FIG. 6 is a flowchart illustrating a prediction model generation method of the storm surge prediction system according to the present invention.
7 is a flowchart illustrating a predictive model learning method of the storm surge prediction system according to the present invention.
8 and 9 are diagrams illustrating input variables and output values used in the storm surge prediction system according to the present invention.
FIG. 10 is a diagram showing the maximum region height value according to the region output by the prediction model.
11 is a graph comparing results obtained by applying the predictive model of the present invention with actual observed results.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention, and the manner of achieving them, will be apparent from and elucidated with reference to the embodiments described hereinafter in conjunction with the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. To fully disclose the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

1 is a schematic diagram of a prediction method of a storm surge prediction system according to the present invention.

As shown in FIG. 1, in order to predict a storm surge, a numerical model is generated based on the virtual storm scenario S1 (S2). Or the weather data S3 for the past storm surge and observation and experiment data S4 are constructed.

The generated weather variables of the typhoon are stored in the database, and the data of the water level, blue, tide and algae as a result of the numerical model are stored in the database as the sea parameters of the typhoon.

The storm surge system of the present invention generates a predictive model by correlating the weather variables and the sea parameters of the typhoon.

The numerical model is used to solve a mathematically defined equation by a numerical calculation process. However, the prediction model of the present invention does not directly calculate mathematically defined equations but statistically analyzes the correlation between input and output variables, And to derive the result.

Based on the observed data (S3) and the experimental data (S4), the actually measured weather data is applied to the predicted model, and compared with the observed data for the actually generated typhoon and the result of the predicted model The prediction model is verified (S6).

At this time, when the prediction model generates a result within a predetermined error range, that is, when the error rate according to the result prediction meets a predetermined condition, it is used as a prediction model for the storm surge.

The meteorological data is input to the currently generated typhoon (S7), and a prediction result is generated based on the prediction model (S8). The prediction result can be used for generating a weather forecast or for a disaster warning.

At this time, JPM-OS (Joint Probability Method-Optimal Sampling), Empirical Simulation Technique (EST), and Frequency Analysis (FA) techniques are used for the weather storm scenario described above.

JPM-OS (Joint Probability Method-Optimal Sampling) creates a new probability distribution by probabilistically combining the angular distributions of the hurricane parameters and creates an optimal scenario based on this. The typhoon parameter is the weather variable of the typhoon. It includes latitude, longitude, central pressure, speed, direction, and maximum wind direction. Empirical simulation technique (EST) is to generate scenarios statistically without assuming a certain probability distribution based on past hurricane parameters. In addition, frequency analysis (FA) is one of the most common statistical techniques, and assumes a certain probability distribution of hurricane parameters that occurred in the past, and creates scenarios based on this. Among these, JPM-OS is the most mature method, and the easiest and most common method is FA.

Also, for the numerical model, STWAVE + ADCIRC, SWAN + ADCIRC, or FVCOM may be used.

STWAVE + ADCIRC stands for a coupled model of the STWAVE model for calculating the wave and the ADCIRC model for calculating the water level. Where STWAVE is a steady state spectral wave model. Similarly for SWAN + ADCIRC, SWAN is the blue model and ADCIRC is the water level model. Here, SWAN is a spectrum wave model similar to STWAVE, but difference from STWAVE differs in that time variation can be considered. FVCOM is a water level calculation model and it is possible to calculate three dimensions (surface xy direction and direction z direction). In addition to water level calculation, it is possible to calculate tide, algae and flow.

At this time, the generation of the virtual typhoon scenario is a step of determining the course, direction, and size of the typhoon, and latitude, longitude, central pressure, direction of travel, speed of travel, and maximum wind radius are determined by the weather parameters .

The meteorological variables of the typhoon generated by the virtual typhoon scenario are used as inputs to the numerical model. As a result of the numerical model, water level and waves are generated. In addition, tides and tides can be output as a result.

2 is a diagram illustrating a configuration of a storm surge prediction system according to the present invention.

2, the storm surge prediction system according to the present invention includes a database 140, a data unit 150, an output unit 160, a prediction model generation unit 110, a prediction result generation unit 130, And a control unit 120 for controlling overall operation.

The database 140 stores the observation data 143, and also stores data for the virtual scenario and the numerical model. In addition, the database 140 stores new weather data to be collected.

At this time, as described above, the weather data 141 of the typhoon by the hypothetical scenario is stored in the database 140, and the sea water variable 142 of the typhoon by the numerical model is stored. The weather variables of the typhoon are latitude, longitude, center pressure, direction, speed of progress, maximum wind radius, and the sea level variables of the typhoon are water level and waves, tide and algae.

In addition, the database 140 stores data on a prediction model generated by the storm surge prediction system.

The data unit 150 stores data generated, measured, or operated during operation of the prediction model generation unit 110 and the prediction result generation unit 130. Also, data by the control unit 120 is stored, and data referred to in the prediction model generation is stored.

The output unit 160 outputs the operation state of the storm surge prediction system to the screen, outputs the data stored in the database 140, outputs the prediction result generated by the prediction model corresponding to the input vapor data, Information on the generated prediction model can also be output.

The output unit 160 includes at least one display means for displaying characters, numbers, or images. The output unit 160 also includes a speaker for outputting a predetermined warning sound or effect sound. The output unit 160 constructs a GUI-based screen and outputs the data and prediction results of the database 140.

The prediction model generation unit 110 generates a prediction model for the storm surge based on the weather variables 141 and the marine variables 142 stored in the database 140.

The predictive model generation unit 110 repeatedly learns based on data stored in the database 140 using a neural network, generates a predictive model, tests the selected predictive model, and selects an optimal predictive model based on the result.

The predictive model generation unit 110 repeatedly changes variables, performs learning by changing the number of neurons, divides existing data into various methods, and applies the predictive model to a prediction model.

In addition, when new data is generated, the predictive model generation unit 110 applies new data to update the previously generated predictive model or generates a new predictive model. The prediction model thus generated is used to predict the next storm surge.

The prediction result generation unit 130 inputs the weather data to the prediction model generated by the prediction model generation unit 110, and generates the prediction of the storm surge as a prediction result.

Based on the weather data, the prediction result generation unit 130 can generate a prediction result according to a change in time based on the weather data, and can also generate a prediction result by predicting a maximum resolution according to a location. That is, when the typhoon occurs, the prediction result generation unit 130 generates the prediction result based on the prediction model generated by the prediction model generation unit 130, and inputs the prediction result to the control unit 120.

The control unit 120 controls the input and output of data in the system, and applies control commands such that each unit performs a predetermined operation. At this time, the control unit 120 receives commands through a separate input unit, and can transmit and receive data through the communication module.

The control unit 120 controls the data to be stored in the database 140 and the data unit 150 and calls the stored data to be input to the prediction model generation unit 110 and the prediction result generation unit 130.

The control unit 120 applies a control command to the predictive model generation unit 110 for at least one of the predictive model generation command, the predictive model generation command, To generate a new prediction model or to update a previously generated prediction model.

When a prediction command for a storm surge is inputted, the control unit 120 applies the input weather data to the prediction result generation unit 130 and controls the prediction result generation unit 130 to operate.

The control unit 120 stores the prediction result based on the prediction model from the prediction result generation unit 130 in the data unit 150 or the database 140 and outputs the result through the output unit 160. [ The control unit 110 causes the prediction result to be configured as a GUI-based screen and output through the output unit.

Also, the controller 120 can transmit the prediction result to the external device through the communication module, and can transmit the prediction result or the warning to the designated terminal or a disaster warning system connected to the system.

FIG. 3 is a diagram for explaining generation of a prediction model of the storm surge prediction system of the present invention, and FIG. 4 is a graph showing learning and verification in prediction model learning of the present invention.

The prediction model generation unit 110 generates a prediction model based on the typhoon weather variables and the sea parameters stored in the database 140. That is, the prediction model generation unit 110 analyzes the weather and sea parameters of the typhoon generated by the existing virtual scenario and the numerical model, and generates a prediction model by setting a function according to the relationship.

The predictive model generation unit 110 can generate a predictive model in a short time by analyzing the result data by the virtual scenario and the numerical model and statistically construct the predictive model so that the predictive model can be changed even if new data is generated.

The predictive model generation unit 110 includes a model construction unit 111 and a learning unit 112.

As shown in FIG. 3, the model building unit 111 constructs an initial prediction model so that a result is obtained through calculation of a multi-layer structure with respect to an input.

The learning unit 112 divides the data stored in the database 140 and applies it to the initial prediction model to analyze the accuracy of the result. At this time, the learning unit 112 not only changes and applies the data, but also modifies and changes the prediction model so as to approximate the resultant value of the stored data while changing the variable value of the layer of the initial prediction model.

The learning unit 112 changes the prediction model so that the result satisfying the condition is output while adjusting the data and the variable more than the predetermined number of times. Since the learning unit 112 uses the past data, the learning unit 112 generates a prediction model by optimizing the variable values included in the layer of the intermediate process, based on the result data on the input that has already been determined.

Further, the learning unit 112 tests whether or not the resultant value satisfies the predetermined condition with respect to the adjusted and predicted model. The learning unit 112 sets the prediction model in which the best result is generated with respect to the result as the final prediction model. The prediction model thus determined is used by the prediction result generation unit 130 to generate an actual prediction result.

At this time, the model building unit 111 constructs a prediction model in a multilayer feedforward manner.

The model construction unit 111 sets the input layer 210 to six input variables, latitude, longitude, center pressure, speed, direction of travel, and maximum wind radius, and outputs the time series storm model or a static model is output.

The model building unit 111 constructs a hidden layer 220 which is a layer to be changed in the course of learning between the input layer 210 and the output layer 230 so that the accuracy of prediction is improved through learning.

The model construction unit 111 sets a predetermined number of nerve cells to the hidden layer 220 and sets the number of neurons to a variable input by applying the weighting W and the bias b and b1, And the intermediate value is input to the output layer 230 to output the resultant value. The model construction unit 111 performs the calculation by applying the weight W2 and the bias b2 to the intermediate layer in the hidden layer 220 as well as the output layer 230. [

At this time, the model construction unit 111 sets the number of neurons of the hidden layer 220 to 16 in the initial prediction model. The number of neurons of the hidden layer 220 is changed depending on whether or not the result in the output layer 230 satisfies a correlation coefficient that is an allowable criterion. In the learning process of the learning unit 112, the number of neurons can be changed from a maximum of 16 to a maximum of 25.

Since the model construction unit 111 uses six variables in the input layer 210, it is possible to set the number of nerves by applying 2 to 3 times of the variable.

The learning unit 112 learns prediction models according to the Levenberg-Marquardt (backpropagation) algorithm. At this time, the learning unit 112 repeats the learning while dividing the stored data, and repeats the learning while changing the number of neurons of the hidden layer 220.

The learning unit 112 terminates the learning when the result satisfies the correlation coefficient, which is the allowable criterion, and changes the number of neural numbers when the result does not satisfy the correlation coefficient. Repeat.

At this time, if the number of neurons does not satisfy the acceptance criterion, the learning unit 112 re-learns one by one while increasing the number of neurons. In the process, the learning unit 112 repeats the data division.

The learning unit 112 adjusts the tolerance level downward if the upper limit of the nerve number 25 does not satisfy the tolerance standard, and the nerve number is initialized and starts again from 16.

The learning unit 112 divides and uses the data. The data division is repeated several times, and the learning is repeated until the result satisfies the correlation coefficient.

For example, the learning unit 112 allows 70% of the data to be used for learning, 15% to be used for the verification, and 15 to be used for the test. At this time, the data used in learning, verification, and testing are not overlapped so that the data is divided to ensure mutual independence.

At this time, if the data is used for 100% learning, there is no data to verify the learned prediction model. Therefore, the data is divided and used for learning as described above. The percentage of data splitting described above can be changed, for example.

The learning unit 112 uses a randomly selected method for data partitioning. If the data is selected by a specific rule rather than an arbitrary method, the prediction model can also be biased because the artificial tendency is strongly influenced by the rule.

The learning unit 112 performs 20 arbitrary divisions, and each divided data is used for learning and correction and verification, and the learned network is subjected to the acceptance criterion. If the tolerance criterion is not satisfied, arbitrary division is repeatedly performed up to 20 times.

At this time, the learning unit 112 performs learning by performing 20 data division on the number of nerves of the hidden layer 220, and performs learning by performing 20 data division on the changed number of neurons when the number of neurons is changed.

That is, when the number of neurons sequentially increases from 16 to 25, the learning unit 112 performs a total of 10 learning operations 20 times each, so that the learning unit 112 performs 200 learning operations. When the number of neurons or the number of data division is changed, the number of learning times of the learning unit 112 is also changed.

The learning unit 112 performs learning from the output layer 230 to the hidden layer 22 in the reverse order to the initial prediction model configured by the feed forward (forward feed).

The learning unit 112 repeats a series of operations for multiplying the weights and weights of the hidden layer 220 and the output layer 230 with respect to the input variable of the input layer 210 several times, .

Since the output value of the input variable of the input layer is a predetermined value and the output calculated by the predictive model differs from the output that is present as the initially set output, i.e., data, the learning unit 112 minimizes the error of the output Repeat the learning process. The learning unit 112 uses the back propagation algorithm to update the weights and the bins of the output layer 230 as the learning is repeated and the weights and the bins of the hidden layer 220 to be updated next.

The learning unit 112 repeats learning until a maximum of 200 epochs as the number of neurons and data are repeated.

The learning unit 112 repeats the learning process until the correlation coefficient between the calculated output and the target output is greater than the initial correlation coefficient. For example, the learning unit 112 sets an initial correlation coefficient of 0.995 in the initial prediction model, and it is compared with the initial correlation coefficient by performing 200 iterations at each time. As described above, since the number of neurons is 10 and the number of data is 20, it is 200 times. At this time, if the correlation coefficient calculated up to the 200th is smaller than the initial correlation coefficient, the learning unit 112 resets the initial correlation coefficient by -0.001 and performs the learning process again from the beginning.

The learning unit 112 stores the prediction models satisfying the target tolerance criteria in each learning process every hour or region. The temporal prediction model is a dynamic model for predicting time series, and the regional prediction model is a static model for calculating only the maximum value.

In addition, if the mean square error of the result is not significantly changed within 200 times, the learning unit 112 stops without repeating the learning. At this time, the learning unit 112 verifies the current prediction model through data division validation and test data.

The learning unit 112 compares the result of the verification data with the result of the learning data to determine whether to continue or stop the learning. The learning unit 112 stops the learning based on the result of the verification data.

At this time, as described above, the learning unit 112 stops the learning when the mean square error of the result is not largely changed, but proceeds to the learning of the basic number six times based on the stopping time.

For example, as shown in FIG. 4, in the 27th learning, when it is determined that the range of the mean square error does not change greatly and the learning is stopped, the minimum number of times is further increased to 6 times, From the verification data for the 27th learning, the mean square error for the verification result is calculated for the verification data for the 33rd learning.

In other words, the basic number of times for verification is 6, and the learning is performed at least six times based on the time when the learning is stopped, and then the verification is performed using the divided verification data for the learning.

From the results for the 27th to 33rd validation data, if any of the errors for the 27th validation to the 33rd validation is not lower than the error for the 27th validation, The result of the 27th learning is set to the highest performance value.

Accordingly, the learning unit 112 stops the learning and sets the prediction model for the 27th learning as the optimal prediction model.

On the other hand, when the result for the 29th verification data has a lower error, 6 times verification is performed based on the 29th verification data. That is, the highest performance value is set based on the results of the 29th to 35th verification data.

At this time, the error of the result by the learning data is lower than the error of the result of the verification data. For example, when the learning data is 70, the verification result is 15, so the result of the learning is based on 70 data, and the result of the verification is based on 15 data. Therefore, The error is lower than the mean square error of the result of the verification, that is, the performance is good.

At this time, the learning unit 112 may perform additional verification based on data other than the database when verifying the prediction model. You can also test predictive models based on test data.

When the optimal prediction model is set by the learning unit 112, the control unit 120 predicts a storm surge based on the prediction model. When a typhoon prediction command is inputted, the control unit 120 inputs a control command to the prediction result generation unit 130, and predicts the typhoon based on the prediction model set by the learning unit 112. [

Based on the prediction model, the prediction result generation unit 130 outputs the time series of the typhoon or the highest region of the typhoon as the prediction result using the input weather data.

At this time, the prediction result generation unit 130 performs time series filtering and filtering on the result at the time of prediction. The filtering only applies to the dynamic model, which is a time series prediction.

There is a problem in that an appropriate value can not be selected when the prediction result has large variability because the correlation value with time before or after the calculation is not considered. Accordingly, the prediction result generating unit 130 excludes a value having large variability through filtering and generates a prediction result. At this time, the prediction result generation unit 130 may perform filtering using the moving average and the Savitzky-Golay filter.

The control unit 110 outputs the output prediction result through the output unit 160.

5 is a flowchart illustrating an operation method of a storm surge prediction system according to the present invention.

5, the control unit 110 inputs a control command to the prediction model generation unit 110 when a command based on the prediction model generation is input.

The predictive model generation unit 110 receives the weather variables and the sea parameters of the typhoon stored in the database 140, and generates a predictive model (S310).

At this time, the model construction unit 111 of the predictive model generation unit 110 constructs a predictive model of a multi-layer structure of the input layer 210, the hidden layer 220, and the output layer 230, (S320).

The model building unit 111 uses latitude, longitude, central pressure, traveling direction, traveling speed, and maximum wind radius as input variables of the typhoon. At this time, since the number of input variables is six, the model construction unit 111 sets the number of nerves by 2-4 times. For example, the model construction unit 111 sets the number of nerves on the basis of 16, but changes it to 25.

The learning unit 112 divides the data stored in the database 140 in step S330 and performs learning on the initial prediction model set by the model building unit 111 using the learning data in step S340.

If the error satisfies the target value during the learning (S350), that is, if the error of the result satisfies the correlation coefficient, which is the allowable standard, the learning unit 112 ends the learning .

If the error of the result does not satisfy the target value, the learning unit 112 performs the learning with a maximum of 200 times while repeating the learning through the data division and the change of the nerve number as described above. In this case, if the error of the result does not satisfy the target value even through 200 times of learning, the target value is changed and the learning is performed again.

When the learning is completed, the learning unit 112 sets a prediction model indicating the best performance, that is, a prediction model having a low error with respect to the result as the final prediction model (S360).

The control unit 110 stores the information on the final prediction model in the data unit 150 and the database 140 when the prediction model generation unit 110 sets the final prediction model.

The control unit 110 receives the weather data and inputs a prediction command for the typhoon warning (S370), and applies the control command to the prediction result generation unit 130. [

The prediction result generation unit 130 generates a prediction result for the input weather data using the prediction model (S380). The prediction result generation unit 130 outputs a tidal-wave solution according to the time of the typhoon or a tidal-wave solution solution corresponding to the region as a prediction result, depending on whether it is a dynamic model or a fixed model.

The prediction result generation unit 130 filters a value with a high degree of variability because the correlation between the predicted results before and after the time is not taken into account with respect to the time dependence.

The control unit 130 outputs the prediction result data generated from the prediction result generation unit 130 through the output unit 160 (S390).

When new data is collected, the controller 120 stores the new data in the database 140, and controls the prediction model generation unit 110 to update the existing prediction model or to generate a new prediction model.

FIG. 6 is a flowchart illustrating a prediction model generation method of the storm surge prediction system according to the present invention.

6, when an initial prediction model is generated by the model building unit 111 (S410), the learning unit 112 causes the learning to be performed based on the initially set number of nerves 16 (S420) .

At this time, the learning unit 112 divides the data stored in the database 140 (S430). The learning unit 112 is divided into training data, verification data, and test data, and is divided by 70%, 15%, and 15% using the arbitrary division method. At this time, the division ratio can be changed.

The learning unit 112 learns the prediction model using the number of neurons and the data segmented by learning data (S440).

An output is determined by repeating a series of operations of multiplying weights and bias times in the output layer 230 of the initial predictive model and the hidden layer 220 in the process of repeated learning, The weight and the bias are updated by the back propagation algorithm, and the prediction model is learned accordingly. At this time, the learning unit 112 performs learning on learning data, for example, time-based learning data or region-based learning data.

The learning unit 112 determines whether the error of the result satisfies the target value (S450).

If the error of the result does not satisfy the target value, the learning unit 112 determines whether the number of data division has been reached (S60). At this time, data division is performed 20 times in total while maintaining the number of nerves, and it is judged whether or not learning of each learning data is completed.

In the case where the number of times of data partitioning is not reached, for example, data partitioning is performed 15 times, data partitioning is performed (S430), and learning is performed on the newly divided learning data (S440).

When the learning is completed through the division of 20 data, the learning unit 112 determines whether the current number of neurons is the maximum number of neurons (S470).

If the number of neurons is 16, the number of neurons is increased by 1 to 17 (S420), and data partitioning (S430) and prediction model learning (S440) are repeated.

If the target value is not satisfied even after a total of 200 learning operations are performed up to the maximum number of neural nets after repeating the above process, the learning unit 112 changes the target value, which is the correlation coefficient for the error (S480) The learning is repeated again from the nerve number 16 (S420 to S480).

On the other hand, if the target value is satisfied during the learning before 200 learning operations, that is, the learning unit 112 determines that the target value is satisfied when the error between the result and the target error is less than the set value, Lt; / RTI >

The learning unit 112 verifies the predictive model using the verification data and the test data of the learning data with respect to the predictive model having the highest performance, i.e., the smallest error (S490). The verification result is compared to set the final prediction model (S500).

7 is a flowchart illustrating a predictive model learning method of the storm surge prediction system according to the present invention.

Hereinafter, the case where the learning is terminated by satisfying the target value in the learning process will be described. As shown in FIG. 7, the prediction model is learned based on the data segmented learning data (S550).

The learning unit 112 suspends the learning when the result value is not varied but is within a certain range (S560), that is, when the deviation of the result is small.

At this time, the learning unit 112 performs additional learning for the basic number of times based on the point of time when the deviation of the result is determined to be small.

For example, as shown in FIG. 4, the learning is performed for the 33rd time by further performing the basic number of times 6 times in the 27th learning. The learning unit performs learning based on the verification data for the 27th to 33rd learning (S570), and performs verification of the learning result based on the learning data.

At this time, the verification data is separate from the learning data, and is divided when the data is divided.

In other words, if the learning stoppage is judged from the result of the learning data in the 27th data partition, the learning is performed based on the verification data of the 27th learning data, and the learning is performed on the 33th verification data.

Based on the result of the 27th verification data, it is judged whether the error value of the result in the verification data up to the 33th step is decreased (S580). If the result of the 27th verification data has the lowest error value, As the optimal result.

At this time, the learning unit 112 may perform additional verification based on the test data and determine an optimal result according to the result. In addition, revalidation can be performed based on data other than the data stored in the database 140.

On the other hand, if the result of the verification data after the 27th time is lower, the sixth learning is further performed based on the corresponding degree (S590).

The results of the learning on the verification data are compared again to determine the optimum result and the process is repeated (S570 to S590).

When the optimal result for the minimum error is selected, the learning unit 112 ends the learning (S600). The learning unit 112 sets the prediction model for the optimal result as the final prediction model when the error of the learning result is the minimum (S610).

8 and 9 are diagrams illustrating input variables and output values used in the storm surge prediction system according to the present invention. Figure 8 relates to a dynamic model that shows a change over time.

As shown in FIG. 8, as a result of the hypothetical scenario, the weather parameter, latitude, longitude, center pressure, direction, progress speed, and maximum wind radius of the typhoon are used as the input parameter A1 in the storm surge prediction system. The result (Target) (B1) value is used as wave or wave.

FIG. 9 relates to a static model including a maximum value according to a region.

As shown in FIG. 9, the input variable (A2) for each typhoon and the maximum harmonic load (B2), which is a result of the input variable, are calculated for a typhoon generated in a specific region. The present invention collects such data by region and generates a prediction model.

The predictive model generation unit 110 repeatedly learns based on the input variables A1 and A2 and the resultant values B1 and B2 shown in FIGS. 8 and 9 to calculate the weights of the hidden layer 220 and the output layer 230 And the prediction model is generated by varying the convenience.

FIG. 10 is a diagram showing the maximum sea level elevation by region. As shown in FIG. 10, in the case of the static model, it is possible to output the maximum resolution for each region as a prediction result.

11 is a graph comparing results obtained by applying the predictive model of the present invention with actual observed results. As shown in FIG. 11A, when the past data is input to the predictive model of the present invention and the result is derived, it is found that the resultant value and the predicted value appear similar when compared with the past observed results have. As a result of comparing the predicted result with the actual result as shown in FIG. 11B, the correlation coefficient for the error is 0.994, which is a very low error rate.

Therefore, according to the present invention, a prediction model is built inversely from input values and output values by using a weather parameter and a sea parameter of a typhoon, which is a result value derived from a hypothetical scenario and a numerical model, It is possible to generate a prediction model with a small number of pixels.

Accordingly, the present invention can quickly and accurately generate a prediction model without an excessive computation process, and when new data is added, the prediction model is modified by rebounding the new data, thereby applying a prediction model suitable for the changing environment, The prediction result can be outputted.

The present invention is not necessarily limited to these embodiments, as all the constituent elements constituting the embodiment of the present invention are described as being combined and operated in one. Within the scope of the present invention, depending on the embodiment, all of the components may operate selectively in combination with one or more.

The foregoing description is merely illustrative of the technical idea of the present invention and various changes and modifications may be made by those skilled in the art without departing from the essential characteristics of the present invention.

110: prediction model generation unit 111: model construction unit
112: learning unit 120:
130: prediction result generation unit 140: database (DB)
150: data portion 160: output portion
210: input layer 220: hidden layer
230: output layer

Claims (10)

A database (DB) in which weather variables of the typhoon, sea variables, and observation data are stored;
A prediction model generation unit for generating an initial prediction model using a weather parameter and a sea parameter of the typhoon stored in the database and generating a prediction model through iterative learning on the initial prediction model;
A prediction result generation unit for generating a prediction result for the input weather data using the prediction model generated from the prediction model generation unit; And
And an output unit for outputting the prediction result,
Wherein the prediction model generation unit includes: a model construction unit that generates the initial prediction model in a feedforward manner of a multi-layer structure including an input layer, a hidden layer, and an output layer; And
And an learning unit for generating the predictive model by modifying the initial predictive model by updating weight values and bias values of the initial predictive models inversely from the output value so that the error of the result value of the initial predictive model is reduced, Prediction system. The method according to claim 1,
Wherein the predictive model generating unit constructs a neural network and generates the predictive model through repetitive learning of the initial predictive model while changing the number of neural networks. The method according to claim 1,
Wherein the predictive model generating unit updates the previously generated predictive model by applying the new data when new data is generated or generates a new predictive model. The method according to claim 1,
The weather variables of the typhoon are latitude, longitude, center pressure, direction of progress, speed of progress,
Wherein the sea level variable of the typhoon is any one of a water level, waves, tides and tides, tide, and algae. The method according to claim 1,
Communication module; And
Further comprising a control unit for controlling input / output of data,
Wherein the control unit controls the prediction unit to output the prediction result through the output unit and transmits the predicted result to the external terminal or the emergency alert system through the communication module. The method according to claim 1,
Wherein the model building unit sets the number of nerves so that the number of input variables of the input layer is two to three times the number of input variables of the hidden layer,
Wherein the initial prediction model is generated by applying weighting (W) and bias (b) to the hidden layer and the output layer, respectively. The method according to claim 1,
Wherein the learning unit divides the data stored in the database into learning data, verification data, and test data in an arbitrary division manner according to a predetermined ratio, and uses the divided data for learning and verification,
Wherein the learning data, the verification data, and the test data are kept redundant and independent. The method according to claim 1,
The learning unit terminates the learning when the resultant value satisfies the target value during learning, changes the number of neural numbers when the result does not satisfy the correlation coefficient, repeats the data division and repeats the learning Emergency storm tsunami prediction system. The method of claim 7,
Wherein the learning unit repeatedly learns the prediction model based on the learning data,
Wherein when the result obtained from the learning data satisfies the target value, the verification is performed on the prediction model learned by the learning data using the verification data or the test data for the learning data. system. Generating an initial prediction model of the multi-layered structure using the weather and sea parameters of the typhoon;
Dividing the data into data and modifying the weight and the convenience for each layer of the initial prediction model through iterative learning on the initial prediction model;
Verifying the modified prediction model using the verification data when the result of the prediction model modified by learning satisfies the target value, and setting the final prediction model;
And outputting the prediction result by applying the prediction model when the weather data is inputted.

Images