時間序列預測有很高的實用價值。隨預測序列長度增長，預測難度也會越來越高，同時附加價值也愈高，因此提升預測序列的長度是非常顯要的研究議題。在許多實際應用當中，如銷售量預測、用電量預測等等，其目標序列都呈現非常明顯的週期特徵，若能有效利用這個特徵，我們可以提升預測遙遠未來的信心水準，進而增加預測模型實際應用價值。 由於近年深度學習方法在許多研究領域上的成功，我們結合了多個深度序列模型並提出了一個encoder-decoder架構的深度網路模型，希望透過融合不同序列模型所提取的多樣時序特徵以提升模型的綜合時序建模能力。 基於對幾個現實序列的觀察，我們發現除了短期的週期特徵以外，這些真實序列也表現出了難以一眼察覺的長期週期特徵或趨勢，這些長期的序列特徵之時間跨度可以長至超越輸入序列長度，使模型無法單從輸入序列提供之線索提取這些隱藏的長期特徵。我們提出了一種日期時間編碼方式，目的是為模型提供更宏觀的時間資訊，以幫助模型提升長期預測準確度。 由於深度網路先天的非線性特性，導致我們的模型對於序列的局部尺度不夠敏感，準確度不如預期。為了解決這個問題，我們參考傳統方法的自迴歸模型，增加了一個與模型並行的線性單元，該單元將歷史觀測經線性映射後產生目標預測序列，並保證預測序列內時間步之間的獨立性以維持模型的平行化特性，最後將此單元的輸出與模型相加以得到更準確的預測結果。 經過實驗，我們提出的架構在幾個現實時間序列資料集上都取得了不錯的成績，並超越了幾個參測基準方法。我們提出的兩大組件在經過消融實驗後證實能在特定情況下有效提升模型的準確度。 ;Deep learning methods has shown superior performance in long-term time series forecasting. For this reason, we propose a deep encoder-decoder neutral network. The model combines multiple deep sequence models, under the hypothesis that integrating diverse temporal features extracted from different sequence models can improve the comprehensive time series modeling capability of the model. Upon analyzing some real-world time series, we found that many of them exhibit short-term periodic features plus long-term periodic features or trends that are not easily discernible at a glance. To model these long-term features, we propose a date-time encoding method to provide proposed model with macro time information to help improve long-term prediction accuracy. Due to the inherent non-linearity of deep neural networks, our model may not be sensitive enough to the local scale of the input sequence, leading to non-satisfactory performance. To address this issue, we introduce a parallel linear unit to the model. This unit linearly maps the historical observations to generate the target prediction sequence. Finally, the output of this unit is added to the output of the deep model for more accurate prediction results. Through experiments, our proposed architecture achieves good results on several real-world time series datasets and outperform several baseline methods. Our two proposed components are proven to effectively improve the accuracy of the model under specific conditions after ablation experiments.
