一個基於深度學習的動態資料驅動應用系統-以天氣預測為例

Abstract

隨著巨量資料趨勢崛起，數量龐大、種類繁多、且具即時性的動態資料日益增加，在資料快速變動的環境中，要達到準確且有效率的資訊預測與評估，成為一個相當困難的挑戰。深度學習是一個新興的特徵類機器學習方法，以多層次類神經網路的方式進行處理，透過抽象層的特徵表示來化簡維度。但是使用深度學習方法時，常會有輸入資料維度過於龐大、無法在動態環境中執行等問題。在本研究中，將深入探討如何應用動態資料驅動的概念，從大量且不同組合所產生的時間序列資料中，找到各種資料與預測目標之間的關聯性，利用關聯分析(Association Analysis)、序列分析(Sequence Analysis)、深度學習(Deep Learning)等方法來設計一個基於深度學習的動態資料驅動應用系統。本研究將以氣象資料為例，相較於先前的研究，此系統的平均預測錯誤率改進了87%。

With the advent of the big data era, dynamic real-time data have increased in both volume and varieties. It is a difficult task to acquire an accurate prediction with respect to rapidly changing data. The Deep Learning (DL) is one of the major approaches of machine learning for feature extraction. It attempts to model high-level abstractions and dimension reduction in data by using multiple processing layers. However, some of the common issues may occur during the implementation process of deep learning, such as: input data having over-complicated dimension, and unable to execute in a dynamic environment. Therefore, it will be helpful if we combine dynamic data-driven concept with DL methods to obtain the dynamic data correlation or relationship between prediction results and actual data in a dynamic environment. This thesis applies the concept of dynamic data-driven to obtain the correlation or relationship between the prediction goals and numbers of different combination results. The methods of association analysis, sequence analysis, and DL are applied to design a dynamic data-driven system based on deep learning. Weather data are used in the experiments. Compare to the previous studies, the proposed system improved the average prediction error rate by 87%.