結合序列到序列和多元狀態向量方法發展水庫入庫流量預報

Abstract

臺灣位於海洋和陸地交界，經常遭受到颱風及暴雨侵襲，於過去10年當中，平均每年約受到4個颱風影響。其豐沛的雨量往往會造成水庫水位迅速上升，若無法在短時間內宣洩，勢必會對人民的生命財產造成威脅，因此，有效率的運用水資源及防災預警至關重要。本研究將運用人工智慧(Artificial Intelligence, AI)方法，建置未來72小時之入庫流量預報模式，另有鑑於臺灣地區降雨分布極度不均的特性，本研究模式將納入此一重要降雨特徵，以期更精確反映實際狀況。 本研究以位於臺中市的德基水庫作為研究區域，蒐集2004至2022年颱風及暴雨事件之觀測雨量及水庫入庫流量，以6種人工智慧模式，如支援向量機(Support Vector Machines, SVM)、隨機森林(Random Forests, RF)、長短期記憶網路(Long short term memory network, LSTM)、雙向長短期記憶網路(Bidirectional Long Short-Term Memory Network, BiLSTM)、閘門循環單元(Gated Recurrent Unit, GRU)及序列到序列(Sequence to Sequence, Seq2Seq)預報未來1小時之流量，並以評鑑指標篩選機器學習與深度學習之較佳模式。將預報未來 1 小時較佳的模式介接多步階預報方法(Multi-step forecasting, MSF)，與多元狀態向量方法(Multi-state-vector, MSV)於兩種輸入項(各雨量站雨量及流量和平均雨量及流量)進行預報未來72小時入流量之比較，並探討不同輸入項對於模式預報準確度之影響，以評鑑指標篩選最佳的預報模式。將該預報模式介接降雨預報產品，分析未來實際應用之可行性。 結果表明，在未來1小時預報中，於機器學習及深度學習模式，分別以SVM及以LSTM為基礎的Seq2Seq模式(LSTM-S2S)表現較佳，其中LSTM-S2S於RMSE、MAE、CE及CC值皆優於其他模式，展現其優秀的性能。在未來72小時預報中，以LSTM結合MSV之模式(LSTM-MSV-S2S)為最佳，於測試場次之RMSE值優於其餘模式約20~35%。於輸入項之比較中，該模式之輸入項採用各雨量站雨量及流量較優，其CC值高達0.871，說明此輸入項能夠反映出整體集水區降雨分布情況，有助於提高入流量預測的準確度。為驗證於實際應用之可行性，將此模式介接降雨預報產品，結果顯示，越臨近流量峰值發生時間，降雨預報產品的準確度將會顯著提高，使流量預報結果能夠更接近觀測值，在防災預警上仍然可提供相當重要的資訊。 本研究提出之LSTM-MSV-S2S且輸入項採用各雨量站雨量及流量之模式，可有效提升水庫入庫流量預報之準確度，於後續應用上，結合降雨預報產品，能夠及早取得入庫流量資料，期望未來可供學研、防災單位參考和做為水庫單位水庫操作策略之依據。

Located at the junction of sea and land, Taiwan is frequently hit by typhoons and rainstorms, with an average of 4 typhoons per year in the past 10 years. The abundant rainfall often causes the reservoir level to rise rapidly, and if it cannot be released within a short period, it will definitely pose a threat to the people's lives and properties, therefore, it is critical to efficiently utilize the water resources and to provide disaster prevention and alerts. In this study, an Artificial Intelligence (AI) method is applied to construct a model for forecasting the inflow for the next 72 hours. In addition, in view of the extremely uneven distribution of rainfall in Taiwan, the model will incorporate this important rainfall feature in order to reflect the actual situation more accurately. In this study, the Deji Reservoir in Taichung City was used as the study area, and the observed rainfall and reservoir inflow from typhoons and rainstorms from 2004 to 2022 were collected and analyzed using six AI models to analyze, such as Support Vector Machines (SVM), Random Forests (RF), Long Short-Term Memory Network (LSTM), Bidirectional Long Short-Term Memory Network (BiLSTM), Gated Recurrent Unit (GRU) and Sequence to Sequence (Seq2Seq) , are used to forecast inflow in the next hour, and the better models of machine learning and deep learning are filtered by evaluation metrics. The better model for forecasting inflow in the next hour is combined with the Multi-step forecasting (MSF) method and compared with the Multi-state-vector (MSV) method for forecasting inflow in the next 72 hours with two inputs. The effect of different inputs on the accuracy of the model was also investigated to select the best forecasting model using the evaluation metrics. The optimal model is then combined with the rainfall forecast data to analyze the feasibility of future practical applications. The results show that the machine learning and deep learning models, SVM and LSTM-based Seq2Seq model (LSTM-S2S), perform better in the 1-hour forecast, and LSTM-S2S outperforms the other models in terms of RMSE, MAE, CE, and CC, demonstrating its excellent performance. In the next 72-hour prediction, the LSTM combined with MSV method(LSTM-MSV-S2S) is the best model, and its RMSE value in the test events is approximately 20~35% superior to the other models. In the comparison of the inputs, the inputs of this model with the rainfall at each rainfall station and inflow are better, and its CC value is as high as 0.871, which indicates that this input can reflect the overall rainfall distribution in the catchment, and help to improve the accuracy of the inflow prediction. In order to verify the feasibility of the model for practical application, this model was combined with a rainfall prediction product, and the results showed that the accuracy of the rainfall prediction product would be significantly improved nearer to the time of the peak flow occurred, so that the flow prediction results could be closer to the observed values, which could still provide important information for disaster prevention and early warning. In this study, the proposed LSTM-MSV-S2S with inputs of the rainfall at each rainfall station and inflow can effectively improve the accuracy of reservoir inflow prediction, and in the subsequent application, combined with rainfall prediction products, the inflow data can be obtained as early as possible, and it is expected to be used for the future reference of academic research and disaster prevention departments, and as the basis of reservoir operation strategy for the reservoir units.