Interview

Abstract

Xia Xu Doctor Xia Xu from Beihang University in China, talks to Electronics Letters about the paper ‘Tropical Cyclone Intensity Prediction Based on Recurrent Neural Networks', page 413. My research field is deep learning, and I am applying deep learning methods for marine disaster (such as typhoons) forecasting. In short, my aim is to predict the expected intensity and path of typhoons. Generally, the mechanisms which govern typhoon behaviour are not clear or well specified, which makes typhoon forecasting a “black box” problem. Deep learning is a new technique uniquely skilled at addressing such “black box” problems and is the method our group chose to apply. Researchers usually try to clarify the mechanisms of typhoons to improve their models and provide more accurate forecasts. However, as is mentioned above, although there are many works about typhoon mechanisms, there is no consensus on an optimal model, and this complicates the use and effectiveness of traditional methods. Typhoon intensity forecasting is a well-recognised international problem. Compared with conventional methods, this work is an entirely different approach, which provides a unique and promising perspective on the problem. In this paper, we reported a new prediction model for the typhoon intensity prediction problem. Our method, deep learning, is not novel in and of itself. However, we have created a new network structure which is able to handle the typhoon intensity prediction task. Thus, the novelty in the presented work mainly exists in the idea and its application, as opposed to the specific technology. We think that this exploratory work will provide inspiration for subsequent research efforts. If this idea is expanded and improved to its fullest extent, it may become a new way to handle the typhoon intensity prediction problem. What was previously considered a traditional challenge will disappear. Furthermore, once deep learning is verified as effective in dealing with typhoon intensity prediction problems, it can then be extended to other similar fields, such as storm surge prediction. These tasks can share the same techniques with only relatively minor model alterations. However, we must face many challenges before reaching this point. Deep learning has had great success in image classification, audio recognition, and other engineering fields, but how can deep learning be geared toward our task? This was the most serious challenge addressed in this Letter. This work has validated the feasibility of using deep learning for typhoon intensity prediction and supplied the drive to continue efforts to further improve the accuracy of this approach. In the short term, more advanced deep learning models and more powerful training tricks may appear to handle this problem. However, in a long term, it is unclear whether our idea will outperform traditional methods. Therefore, with our idea as a first attempt at this approach, future works will explore its effectiveness and validate its eventual success. This initial work is an exploratory one, where a simple RNN model is used to test the feasibility of the approach. It can be easily extended by changing to more powerful models, adding new structures and developing new tricks. In this paper, our accuracy is comparable to traditional methods. In the future, we will work to improve the accuracy, and demonstrate any emerging advantages of the prediction model. While the work presented in our Letter will not, objectively speaking, have an immediate impact in the field it sets an important precedent in the application of deep learning to typhoon forecasting. This work is not in the current mainstream of typhoon intensity prediction as there are not many researchers dedicated to using deep learning methods in this manner. However, I am confident this will change with the continued growth and relevance of deep learning in the coming years. It is a powerful tool, and future research will surely demonstrate its potential and validate its use in our chosen application.