深度信念網路於降雨量預測之探討

近年來隨著科技的發展，每個人生活都遍佈著科技的相關應用，表示著周遭環境可取得各式各樣的數據資料。其中氣象資料對於大眾生活更是息息相關，不管從事農業、漁業或服務業都需要依賴準確的氣象資料做預報工作，更有需多研究與政府單位企圖從中尋找有效的預測方法。本研究研究目的將針對臺灣新北市降雨量取用2017年氣候資料建立預測降雨量模型，試圖預測隔年2018年每小時降雨量。本研究採用指數平滑作為前處理方式結合深度信念網路DBN來建構預測降雨量模型，並採用最小二平方支持向量迴歸(LSSVR)與倒傳遞類神經網路(BPNN)作為比較模型，透過均方根誤差(root mean squared error，RMSE)、平均絕對誤差(mean absolute error， MAE)及取最大值模型效能（Model Efficiency ，EF）三項誤差指標作為評比依據。並進一步探討藉由使用去除連續超過10、24小時降雨量或是去除所有為零的降雨量資料方式凸顯雨季資料進行建模，探討該調整方式是否能有助於增加且有效預測具有豐沛降雨量時段，經實驗結果顯示，本研究提出之指數平滑結合深度信念網路DBN預測模型，預測效能上優於相較其他預測模型；且對於訓練預測模型時去除間歇性需求資料上過多的零值，可有效提高預測模型學習特徵之準確率。

關鍵字

降雨量預測；深度學習；深度信念網路；間歇性需求；指數平滑

並列摘要

In recent years with the development of technology, public daily life is surrounded by technology applications, indicating that people can collect all sorts of data without efforts. In which atmospheric data is highly related to public life, no matter for agriculture, for the fishery or for the service industry, good forecasts rely on accurate atmospheric data. There are many studies and concerned authorities try to find an effective way to forecast with accuracy. In this study, we build hourly rainfall forecast models with the 2017 atmospheric factors dataset in New Taipei City, then test the model with the 2018 dataset. We employ exponential smoothing as a preprocessing method, combined with Deep Belief Networks (DBN) to build the model. Then comparison the model with Least Square Support Vector Regression, LSSVR, and Back Propagation Neural Network, BPNN, using root mean squared error (RMSE), mean absolute error (MAE), and Model Efficiency (EF) to evaluate the performance of the model. Furthermore, we build models by eliminating consecutive zeros over 10 hours, 24 hours and all zeros respectively in terms of rainfall in order to highlight rainfall data. Exploring whether these adjustments help increase effective forecasts in times that actually rain. The result shows that proposed exponential smoothing combined with DBN model outperforms other models, and by eliminating excessive zeros in intermittent demand dataset effectively increase model accuracy on feature learning.