Application of ANFIS on flood prediction in main stream of the Yangtze River

對于易受洪災的地區而言,快速而準確的洪水預報非常重要,能夠為洪水預警消息的發布提供更長的先導時間,從而為可能受災地區的人們提供更充足的時間以采取相應的防洪措施或安全轉移。常用的預報模型包括基于物理性模型和基于系統技術模型。盡管物理性模型能對洪水形成的物理過程提供很好的解釋,水文學家并不愿意使用它們,因為模型中參數的率定是比較復雜的。因此,一種基于純數據集的黑箱技術已被廣泛采納。常用的黑箱模型包括線性模型(LR)、自回歸移動平均模型(ARMA)和人工神經網絡模型(ANN)等。在當前的研究中,一個相對新穎的黑箱模型——基于自適應網絡的模糊推理系統(ANFIS)被用來對長江某河段的洪水進行預報。與此同時,一個線性回歸模型(LR)用來作為ANFIS模型的對照。在構建ANFIS中,混合學習算法(即誤差反衍(BP)耦合最小二乘法(LSE))用來訓練模型的參數。此外,為避免出現過度訓練現象,原始數據集基于統計特征值劃分成3個子集:訓練集、測試集和校正集。當對ANFIS模型訓練時,測試集用來幫助控制訓練代數。結果表明,ANFIS的預報效果優于LR模型。分析認為ANFIS能夠提供預報精度是因為其采用了局部擬合技術,通常它會優于LR模型所采用的全局擬合技術。最后,對本研究而言,最適合的ANFIS模型是輸入量為梯形的成員度函數。||As far as a flood-prone region is concerned,a rapid and accurate flood forecasting is especially significant because it can extend the lead time for issuing disaster warnings,thus allowing sufficient time for people in hazardous areas to take appropriate action,such as evacuation.Although they give a deep clairvoyance to physical mechanism of flood forming,conventional conceptual forecasting models are inconvenient for operational hydrologists in practice.Therefore,many called"black box"models based on systems theoretic techniques,such as linear regression (LR),autoregressive moving average (ARMA), and artificial neural network (ANN),are employed.Compared with conceptual models,often they can provide a rapid prediction with an accepted degree of accuracy in view of depending only on data-driven techniques.In the present study,a relative novel black box technology,namely adaptive-network-based fuzzy inference system (ANFIS) in which Takagi and Segeno's rule was adopted,was proposed for streamflow forecasting in the main channel of the Yangtze River.In the meantime,a linear regression (LR) model was used as the benchmark for ANFIS model evaluation.In the ANFIS model,back propagation (BP) learning algorithm and hybrid learning algorithm (Combined BP and least squared error) were applied to the model,respectively.In addition,in order to avoid overfitting of training data,a statistic information-based data partition technique was used to split raw data into three parts:training data,testing data,and validation data.Of them,testing data played a role as early stopping,which helps obtain the optimal training epoch during addressing training data.Results showed that ANFIS model is superior to the LR model,and the optimal model is the ANFIS model with hybrid learning algorithm and trapezoidal membership functions for the present case.A further analysis revealed the powerful capability of ANFIS is due to the local linear approximation technique being employed in ANFIS model,which improve the capturing capacity for training data if the overfitting can be well controlled.