Keywords

1 Introduction

Currently, natural disasters are being researched in a variety of ways. For example, there have been several investigations into the political [1], economic [2], and statistical aspects of natural disasters [3]. Since natural disasters have a large impact on society, many researchers have made substantial efforts to mitigate the damage that can arise from them. There are many phenomena that are classified as natural disasters such as floods, droughts, and earthquakes. In this paper, we only focus on tropical storms, particularly typhoons.

As technology has developed over the years, tropical storms have been analyzed in many ways including satellite imaging and weather radar [4]. Additionally, the prediction of the movement of tropical storms has also been previously considered [5, 6]. Since analyzing weather information is a big data problem, modern methods of weather prediction or simulation take a significant amount of time even if the calculations are done by a super computer.

To address this issue, we propose a novel method that uses a self-organizing incremental neural network (SOINN) that is trained on big data (i.e., the meteorological information). Since SOINN demonstrates good performance with respect to calculation times, the predicted route of a tropical storm can be calculated quickly. Moreover, accuracy is very important to evaluate the target; therefore, we will also calculate the accuracy of the predicted route. Finally, simulated tropical storms will be presented. Throughout this paper, we use an Intel® CoreTM i7 CPU 870 2.93 GHz with 16.0 GB RAM, and the results are calculated using MATLAB.

2 Training Procedure

2.1 Self-organizing Incremental Neural Network

A SOINN, which was originally proposed by Shen and Hasegawa [7], can be trained without initializing the number of nodes and locations. SOINNs have also previously been reported to display robust performances when handling noisy datasets. Figure 1 depicts the SOINN algorithm [8] that is used in this paper. By applying SOINN, it becomes possible to handle the problem even if there is only few data for training. We have a limitation of data, than, SOINN was decided to adapting for prediction model.

Fig. 1.
figure 1

Algorithm of SOINN

Full size image
Fig. 2.
figure 2

Atmosphere map in 2015/05/05 3:00 Am, East Asia.

Full size image

In this paper, we set \( \uplambda = 1000,\,age_{max} = 5000 \). In SOINN, the nodes deleted if it doesn’t updated until \( age_{max} \). It was reported that if \( \uplambda \) is low, SOINN becomes strong at noise and hard to make cluster, and if \( \uplambda \) is high, it becomes weak at noise and easy to make cluster. We hope that data don’t disappear by \( age_{max} \) and become easier to make cluster, the value of \( \uplambda \) and \( age_{max} \) are considered.

2.2 Tropical Storms

Several names can be used to describe tropical storms, such as typhoons, cyclones, and hurricanes. The name that is applied to a tropical storm in question is selected based on where the storm is occurring. For example, if the storm occurs in the Atlantic of Northeast Pacific, the term “hurricane” is used. If the storm occurs in the Northwest Pacific, it is called a “typhoon”. Lastly, the term “cyclone” is used if the storm occurs in the South Pacific or the Indian Ocean [9].

In this paper, we only consider data from typhoons, and the data for these typhoons are presented in Table 1. The location of the typhoon, starting from when it was born and ending when it disappeared and measured in 3-hour intervals, was acquired from the Japan Meteorological Agency (JMA) [10].

Table 1. Classification of typhoon used in learning (T1511 means 11th typhoon which came in 2015)
Full size table

2.3 Atmosphere

Atmospheric conditions are also considered when training the SOINN on typhoon data. For atmospheric analysis, we set the longitude to the x-axis and the latitude to y-axis. We only considered the longitudinal range of 115–155 and the latitudinal range of 20–50 in this paper. We divided the longitude into 4 parts and the latitude into 3 parts, which divides the data area into 12 parts. We then evaluated each area’s atmosphere based on the following scheme:

  1. (1)

    If high pressure dominates in that area, the area is given a value of 1.

  2. (2)

    If low pressure dominates in that area, the area is given a value of 0.

  3. (3)

    If neither high nor low pressure is dominant in that area, the area is given a value of 0.5.

In this study, digitalizing process divided the study area in 12 parts. It ignores specific distribution, then, the error derived by this ignorance would be appeared. It is needed to consider the division which helps to reinforce the performance.

The atmospheric data was acquired from the Japan Weather Association [11] (Fig. 2).

Fig. 3.
figure 3

Example of evaluating Fig. 2.

Full size image

3 Experiment

To verify the effectiveness of the proposed route prediction method, we used the 5 typhoons listed in Table 2. We used all of the landing typhoons that occurred in 2015 and 2016 to train the SOINN, Thus, we selected typhoons T1408 and T1419, which occurred in the desired time interval. For the approaching typhoons, T1617 was selected because it is the most recent data point for the approaching category. Although T1618 is technically the most recent approaching typhoon, its route is similar to those of T1408 and T1419, and therefore we elected to use T1617 and not T1618. Most of the typhoons that fall into the “no effect” category disappeared very quickly. However, T1504 and T1626 occurred in the desired time interval and existed for more than one week. Therefore, they were selected for analysis.

Table 2. Classification of typhoon used for test. (T1408 means 8th typhoon which came in 2014)
Full size table

When we use the typhoon’s location data for learning and testing, the following normalization is applied. \( T_{Latitude} \) denotes the latitude value of typhoon and MAX \( \left( {T_{La\_all} } \right) \) denotes the maximum value in all \( T_{Latitude} \). \( T_{Longitude} \) denotes the longitude value of typhoon and MIN\( \left( {T_{Lo\_all} } \right) \) denotes the minimum value in all \( T_{Longitude} \).

$$ T_{Normalized Latitude} = \left( {T_{Latitude} - MIN\left( {T_{La\_all} } \right)} \right)/\left( {MAX\left( {T_{La\_all} } \right) - MIN\left( {T_{La\_all} } \right)} \right) $$
(1)
$$ T_{Normalized Longitude} = \left( {T_{Longitude} - MIN\left( {T_{Lo\_all} } \right)} \right)/\left( {MAX\left( {T_{Lo\_all} } \right) - MIN\left( {T_{Lo\_all} } \right)} \right) $$
(2)

3.1 Prediction Method

To predict the route of the typhoon, the following algorithm is used.

figure a

By repeating Algorithm 1 N times, we can predict the location of the typhoon after 3 N hours.

3.2 Location Adjustment

Because only 35 typhoons are used for training the algorithm, the amount of data is not enough. Therefore, we applied the Algorithm 2 after the Algorithm 1 before repeating.

figure b

If the predicted location is evaluated too far by step (2) in Algorithm 2, \( \overrightarrow {{x_{n + 3}^{*} }} \) will be changed to follow their original movement. A similar thing happens when predicted location is evaluated too close. In this report, we set \( {\text{r}}1 = 1.5\,{\text{and}}\,{\text{r}}2 = 0.8 \).

3.3 Result

By using Algorithms 1 and 2, we predicted the route of 5 typhoons for testing the performance. From the JMA homepage, we can check the average error for predicting 24 and 48 h ahead [12]. We also estimated the error for the 24- and 48-hour predictions from our algorithm. During this time, we predicted the movement of the storm until it changed to an extratropical low. The following table presents the average distance between the real storm location and the predicted location. According to Table 3, the error of prediction in landing typhoon is bigger than other error. One of the main reason is shortage of data in landing type. We expect that result might be improved with more data.

Table 3. Average error of location of proposed method
Full size table

We also present the average error for different years from the JMA. The JMA uses two super computers-a FUJITSU PRIMEHPC FX100 and a FUJITSU PRIMERGY CX2550M1-and professional datasets that are not available to the public for calculations [13] (Table 4).

Table 4. Average error of location by Japan Meteorological Agency
Full size table

The total average error of our predictions is 105.5 km for 24 h and 250.0 km for 48 h. The SOINN takes only 0.663 s to be trained on the dataset. The specific calculation times are shown in Table 5.

Table 5. Calculation time for prediction
Full size table

Figures 4, 5, 6, 7, 8, 9, 10 and 11 present the actual and predicted movement of the storm. The cross points in Figs. 7, 8, 9, 10 and 11 indicate the real location of the typhoon. The lines that originate from each cross point indicate the predicted route at that point. Due to space limitations, we only present the actual and predicted routes for T1408 at full size. The remaining figures are presented in a smaller size.

Fig. 4.
figure 4

Average error of location by our method

Full size image
Fig. 5.
figure 5

Comparing error of ours with error of Japan Meteorological Agency

Full size image
Fig. 6.
figure 6

Route of T1408

Full size image
Fig. 7.
figure 7

Predicted route of T1408

Full size image
Fig. 8.
figure 8

Route of T1419 in left and predicted route of T1419 in right

Full size image
Fig. 9.
figure 9

Route of T1504 in left and predicted route of T1504 in right

Full size image
Fig. 10.
figure 10

Route of T1617 in left and predicted route of T1617 in right

Full size image
Fig. 11.
figure 11

Route of T1626 in left and predicted route of T1626 in right

Full size image

In Figs. 7, 8, 9, 10 and 11, lines indicate the predicted movement four days into the future from each point. However, we did not calculate the error in the distance prediction two days into the future.

4 Conclusion

In this paper, we used SOINN to predict the routes of tropical storms. Our proposed method shows a close accuracy to the JMA predictions, even though our method takes only a few seconds for the calculations. Furthermore, we only used open tropical storm data from 2015 and 2016. Though we used only open data that we refined slightly, our method indicates a high accuracy and fast calculation speed. This results based on one of the characteristic of SOINN: online learning. There is substantial potential for improvement if larger, more specific meteorological datasets are used. We, therefore, conclude that our method is effective and performs well in predicting the routes of tropical storms.