- 期刊
應用時間序列模型預測SPI比較分析
Comparison and analysis of time series models for SPI predictions
摘要
本文共建立6種預測標準化降雨指標（SPI）的時間序列模型，包含：長短期記憶法（LSTM）、門閥遞迴神經元（GRU）、雙向LSTM（Bi-LSTM）、雙向GRU（Bi-GRU）、支持向量機及整合移動平均自迴歸模型（ARIMA），以均方根誤差（RMSE）作為評估模型準確性。本文假設未來半年（6筆）SPI的變化僅與過去一年（12筆）的SPI有關，以建立模型訓練、驗證及測試資料。LSTM、GRU、Bi-LSTM及Bi-GRU均以5層多神經元LSTM層（或GRU層、Bi-LSTM層、Bi-GRU層）建立模型，SVM以系統內定超參數進行迴歸計算，沒有進行超參數優化，ARIMA皆以一次差分完成模型建立。研究區域為濁水溪沖積扇範圍內共10個雨量站，月雨量資料範圍為1990～2017，各雨量站之SPI多落在±2之間，且不定期發生連續中度乾旱（SPI＜-1）或嚴重乾旱現象，若能預測未來SPI有助於水資源調配。神經網路的計算以繪圖加速器完成，SVM及ARIMA則以CPU完成計算。由於LSTM神經網路的結構設計，會影響LSTM模型的預測能力，本研究的目的為（1）在計算上建立更有計算效率的LSTM神經網路結構，（2）應用ARIMA、LSTM及其他衍生模型建立各雨量站不同時間尺度SPI預測模型，及（3）評估各雨量站在不同時間尺度SPI的最佳模型。研究結果發現，本研究所建立的5層LSTM模型，在計算時間及精確度上，均明顯優於文獻中所建議的單層LSTM模型，且Early Stopping的訓練流程可以大幅降低計算。建立SPI-12預測模型時，ARIMA可作為多數雨量站的SPI預測模型。而GRU所建立西巒SPI-6預測模型之RMSE（0.2924）最大，ARIMA所建立西螺（2）之RMSE（0.0553）最小，對於預測SPI-3及SPI-6而言，LSTM、GRU、Bi-LSTM、Bi-GRU或ARIMA，都可能作為最佳的預測模型，但沒有一個預測模型可以作為唯一建立雨量站SPI的預測模型，其原因是預測模型的選擇與作為建立模型的訓練資料相關。
並列摘要
In this paper, 6 time series models including LSTM, GRU, Bi-LSTM, Bi-GRU, SVM and ARIMA were established for SPI prediction. RMSE is used to assess the accuracy of model predictions. A 5-layer multiple-neuron were established for LSTM, GRU, Bi-LSTM and Bi-GRU. Default super-parameter of SVM was applied for regression without optimization. First order difference was used for ARIMA model. Monthly precipitation data from 1990-2017 of 10 stations in Chou-Shui River Alluvial Fan were used for model training or establishment. Most SPI values of the precipitation stations are between ±2. Moderate drought or even severe drought was found during the research time. A successful SPI prediction model would be helpful for water resources allocation. As the prediction ability of LSTM model can be determined by its neural network structure, the objectives of this research include: (1) development of a better LSTM model for SPI prediction, (2) establishment of SPI prediction models for different time scales of each precipitation station by ARIMA, LSTM and its extension models, and (3) assessment of the best model for different time scales SPI of different precipitation stations. The results of the study found that the 5- layer-LSTM model developed in the research is much better than the one-layer LSTM model from literature review. It was found that application of Early Stopping in LSTM reduced a lot of computation time. Furthermore, ARIMA can be used as the prediction model for most precipitation stations when establishing the SPI-12 prediction model. Comparing all models of this study, the RMSE (0.2924) of XiLuan SPI-6 prediction model established by GRU is the largest, and the RMSE (0.0553) of XiLuo(2) established by ARIMA is the smallest. LSTM, GRU, Bi-LSTM, Bi-GRU or ARIMA could be chosen to establish prediction model for SPI-3 and SPI-6. However, there is no unique model solution for SPI prediction of 10 precipitation stations. The reason is that the selection of prediction model is related to the training data.