演化式類神經網路於水文系統預測之研究

Abstract

智慧型之優選系統應具備自我調適以適應各類型問題之能力，而學習與演化為調適之二種基本型式。一般之優選系統多僅具備其中一種能力，如類神經網路具優異之學習能力，而演化法擁有絕佳之演化能力。傳統類神經網路之建構方式一般係預設固定之網路架構，惟此舉並無法自動地探尋最適合訓練資料之網路模式。而演化法可依生物演化之特點，針對擬求解問題之特性自動地調適網路架構或連結權重。為改善傳統類神經網路優選過程之缺失，本研究提出一嶄新智慧型之演化式類神經網路－「混合編碼演化式類神經網路」，此法以獨創之混合編碼方式，將多層前饋式類神經網路之架構參數（包含輸入變數及隱藏層之神經元）編碼，並以遺傳演算法搭配採用比例共軛梯度法之倒傳遞類神經網路及採用最陡坡降法之線性類神經網路，同時優選網路之架構及各神經元間之最佳連結權重。 本研究建構之混合編碼演化式類神經網路，具下列特色：（一）可自動優選多層前饋式類神經網路之最佳架構；（二）可自動優選出最佳輸入變數之組合；（三）經編碼後之網路架構染色體可於演化過程中進行基因交配運算，並於基因解碼後轉換為不同層數之網路架構；（四）兼具處理線性及非線性優選問題之能力。 由Mackey-Glass混沌時間序列預測之結果，顯示本研究所建構之混合編碼演化式類神經網路兼具效能性、效率性及強健性。另本研究將建構之演化式類神經網路應用於水庫入流量、水庫水位及蒸發量等不同水文系統之預測及推估，結果顯示相對於傳統倒傳遞類神經網路、AR(1)與ARMAX二種時間序列預測模式或Modified Penman蒸發量推估模式，演化式類神經網路之改善百分比均能達到10%以上。

Intelligent optimal systems should have the ability of self-adaptation in order to adjust to various problems. Learning and evolution are two fundemental forms of the adaptation ability. However, the common optimal systems only have one of the two above abilities, for example, artificial neural networks (ANNs) with excellent learning and evolutionary algorithms (EAs) with admirable evolution. The conventional ways of constructing ANNs for a problem generally presume a specific architecture and do not automatically discover network modules appropriate for specific training data. EAs are used to automatically adapt the network architecture or connection weights according to the problem environment without substantial human intervention. To improve on the drawbacks of the conventional optimal process, this study presents a novel intelligent evolutionary artificial neural network, so-called hybrid-encoding evolutionary artificial neural network (HEEANN), for time series forecasting. The HEEANN has a hybrid encoding and optimization procedure, including the genetic algorithm, the scaled conjugate gradient algorithm, and the gradient descent, where the feed-forward ANN architecture (including the inputs and the neurons in hidden layers) and its connection weights of neurons are simultaneously identified and optimized. The proposed HEEANN has the abilities to: (a) automatically optimize architecture of feedforward ANNs; (b) automatically optimize input variables among all possible ones; (c) evolve network architecture with different-length hidden layers by performing crossover operation between ecoded architecture chromosome; (d) automatically deal with linear and non-linear optimization problems based on whether the evolved network architecture has hidden layer when prior information of the data is insufficient. We first explored the performance of the proposed HEEANN for the Mackey-Glass chaotic time series. The performances of the different networks were evaluated. The excellent performance in forecasting of the chaotic series shows that the proposed algorithm concurrently possesses efficiency, effectiveness, and robustness. We further explored the applicability and reliability of the HEEANN in several real hydrological time series. Again, the results indicate the HEEANN is supeior to backpropagation neural network (BPN), AR(1) and ARMAX models, or Modified Penman model with over 10% improvement.