Fuzzy Neural Networks And Genetic Algorithms Applied Research, Real-time Flood Forecasting

Hydrologic forecasting, which is a complex system with strong nonlinear feature, is a very crucial non-structural measure of flood mitigation to adapt nature and mitigate losses, and directly serves for flood defense, reasonable utilization and protection of water resources, construction, operation and management of hydraulic structure and developments of industry and agriculture. Nowadays, whether deterministic hydrologic model or stochastic hydrologic model are founded upon observed data. Due to the restriction of assumptions in model-construction, these models, in a great sense, are a kind of analogy of actual hydrologic laws, and are hard to deal with complex nonlinear relations among hydrologic phenomena and their factors. Based on a summary of domestic and abroad study, this paper tries to establish fuzzy optimizing neural network real time flood forecasting model, which combines Artificial Neural Network (ANN), Fuzzy Sets theory and Genetic Algorithms (GAs). The framework and research fruits are listed as follows:1. Aiming at the fuzziness and complex relations of hydrologic phenomena and their factors, combining excellent knowledge expressing ability of fuzzy mathematics with fine learning ability of ANN, real time forecasting of flood process is implemented using fuzzy optimizing neural network. Using GAs to optimize initial weights of neural network, in a sense, local optimizing problems, which is widely existed in neural network model training, can be overcome. According to variant thoughts for simulating flood process, two forecasting models (Model I and II) are proposed. With variant input factors, model I has two cases. Based on memberships of rain gauging stations attributed to upper and lower reaches, the spatial distribution of rainfall is also discussed.2. Taken Shuangpai reservoir in Hunan province as an example, model I and II are testified, the determination of input parameter studied, result-reasoning carried out, and based on these, structure of forecasting model is finally determined.3. No existing referential formulae are ready to calculate parameters, such as number ofhidden layer nodes in neural network and parameters in GAs, and commonly parameters are determined using testing methods. Sampling method is used to specify optimal parameters of model. Due to the independence of parameters, we change the parameter to be specified and keep other parameters constant to get the optimal value of this parameter. To avoid the randomicity of calculation, we run the models twice, calculate the mean value of network output error and draw its graph. After analyzing the influence of parameters on computing results of model, optimal parameters is finally specified. Using optimized parameters to carryout analogy and forecasting, an effective real time flood forecasting model is explored.