Deep neural network based feature representation for weather forecasting

This paper concentrated on a new application of Deep Neural Network (DNN) approach. The DNN, also widely known as Deep Learning(DL), has been the most popular topic in research community recently. Through the DNN, the original data set can be represented in a new feature space with machine learning algorithms, and intelligence models may have the chance to obtain a better performance in the “learned” feature space. Scientists have achieved encouraging results by employing DNN in some research fields, including Computer Vision, Speech Recognition, Natural Linguistic Programming and Bioinformation Processing. However, as an approach mainly functioned for learning features, DNN is reasonably believed to be a more universal approach: it may have the potential in other data domains and provide better feature spaces for other type of problems. In this paper, we present some initial investigations on applying DNN to deal with the time series problem in meteorology field. In our research, we apply DNN to process the massive weather data involving millions of atmosphere records provided by The Hong Kong Observatory (HKO)¹. The obtained features are employed to predict the weather change in the next 24 hours. The results show that the DNN is able to provide a better feature space for weather data sets, and DNN is also a potential tool for the feature fusion of time series problems.