遺傳演算法(Genetic Algorithm,GA)與模擬退火演算法(Simulated Annealing,SA)的混合應用最早是由Goldberg於1990年提出Boltzmann Tournament Section(BTS)。模擬退火演算法 具有較佳的局部搜尋能力,相對的實數編碼利基遺傳演算法具有較佳的整體搜尋能力,但是最佳解還是不易求得,本研究乃綜合兩者不同演算法所具有的特性,彌補單一演算法的不足,並結合田口(Taguchi, 1950)初始值設定法的應用,進行優選問題之求解,經數學函數測試,本研究所使用的方法收斂情形較佳,且較容易找到全域最佳解。 本研究所採用的模式為「全流域河川不恆定流模式」(賴經都, 1999),模式中為維持對低中水位及高洪水位計算模擬之準確度,因此,水流阻力參數採用可變值,隨水深而變化。本研究將嘗試同時自動率定低中水位及高洪水位阻力參數。本研究方法首先利用田口初始值設定法找出一組初始解後,再經模擬退火實數編碼利基遺傳演算法自動率定最佳阻力參數。研究範圍以淡水河流域之新海橋、河口、中正橋及南湖大橋水位站為邊界點,選用流域重大颱洪事件(納坦颱洪)實測之洪水歷線資料來進行率定;參數率定後,並利用不同颱洪事件(海馬颱洪、艾利颱洪)之實測水位記錄資料來驗證模式之適用性。 研究結果顯示:率定所得的阻力參數,可使模式之計算水位與觀測水位相當吻合。以另一時段觀測資料進行模式驗證時,發現計算水位與觀測水位亦十分接近。可見本研究所提議的方法於「全流域河川不恆定流模式」之阻力參數率定,的確是一客觀且能正確找出全域最佳解的自動化參數率定方法。
Mixed use of Genetic-Algorithm (GA) and Simulated-Annealing (SA) was proposed as Boltzmann Tournament Section (BTS) by Goldberg in 1990. The SA has better local search capability, but GA has better global search capability. Combined features of these two algorithms are aimed to remedy the deficiency of each single algorithm. It is further to integrate Taguchi approach (Taguchi, 1950) for better initial guessed values to accelerate the process of searching. Present method has illustrated its fast convergence over other methods. In this thesis, a river flood simulation model CCCMMOC(Lai, 1999) is adopted for water levels simulation. In this model, the river hydraulic resistance coefficients are unknown to be calibrated and assumed to change with depths of water. Calibrations of the resistance coefficients at the lower and higher water levels are performed separately. In this study, Taguchi initialization method is used first to select the initial gauss solutions. It then employes the Simulated Annealing Real-valued-coding Niche Genetic Algorithm (SARvcNGA) to calibrate the best resistance coefficients. Observed data of Natain Typhoon (2004) flood levels is Tamsui River are used for calibration. Historical observations of Typhoons, Heima (2004) and Aili (2004) are used for verifications of the calibrated river resistance coefficients. Calibration efficiency of present model is demonstrated. Present approach has been illustrated to be efficient for automatic river simulation model calibration.