應用深度學習於具容量限制服務業之需求預測－以自助洗衣業為例

Abstract

具容量限制的服務業與人們的生活息息相關，例如餐廳、洗衣店、電動車的充電樁等。對於該產業而言，準確的預測使用者需求十分重要，因為設置過多的量能會讓資源閒置；而過少的量能會讓想上門消費的客人撲空。過往的研究中，針對AI 輔助具容量限制服務業的研究非常稀少，因此我們希冀提出一個創新的使用者需求預測模型以協助該產業。

本研究提出一個全面的預測平台，涵蓋了兩個模型，分別是時間序列分析模型以及兩階段分類模型。在時間序列分析模型中，我們提出了創新的縱向時間序列資料轉換方式，並結合了循環神經網路，以預測那些已具有歷史資料的營運地點。另外在兩階段預測模型中，我們則採用了新穎的資料索引方法、聚類分析和分類演算法，提供給沒有歷史資料、新設立的營運地點來預測使用者需求。

本研究採用了連鎖洗衣店的資料集來驗證此預測平台的預測效果。實驗結果證實了在連鎖洗衣店的資料集上，時間序列分析模型以及兩階段分類模型都能夠準確地預測使用者需求。且時間序列分析模型的表現遠比傳統的差分整合移動平均自我迴歸模型較佳。另外，透過實驗的過程，我們也確立了應用該預測平台的最佳條件。

The capacitated service industry, such as restaurants, hotels, and laundromats, plays an important role in people's lives. It is important for the industry to accurately predict user demand and prevent resource waste or potential loss of profits. However, there has not been much research on AI assistance for the industry before. With the development of AI, we aim to propose a new user demand forecasting method to help the industry.

In this study, a comprehensive user demand forecasting platform is proposed. The forecasting platform contains two models, namely the time series analysis model and the two-stage classifying model. For the time series analysis model, a novel vertical time series data transformation method and recurrent neural network are applied to predict user demand for locations with historical data. As for the two-stage classifying model, an innovative data indexing method, clustering, and classification are adopted to predict user demand for locations without historical data.

To validate the proposed models, the dataset collected from a laundry chain is employed. From the experimental results, both the proposed time series analysis model and the two-stage classifying model are capable of predicting user demand accurately. Moreover, the performance of the time series analysis model significantly exceeds that of the baseline ARIMA model. In addition, the best conditions for applying the two models are determined through the experiments.