Muti-step-ahead Prediction And Optimization Algorithms For Nonlinear System And Their Applicatons In Hydrologic Forcasting

The prediction and the optimization of nonlinear dynamic systems widely exist in various domains such as weather, hydrology, medicine, electronics and information science. The importance of them should not be disregarded. On the one hand, compared with single-step-ahead prediction, multi-step-ahead prediction, which can provide prediction of many time-steps into the future, is desirable in practice. On the other hand, in the area of nonlinear optimization, multi-modal function optimization is a difficult and hot problem. The phenomena of adaptive optimization in the biology immune system have given us many metaphors. Creatures and natural ecological systems can solve many optimization problems with high degree of complexity to scientists satisfactorily and elegantly. They provide us many new methods to tackle multimodal function optimization problems which are difficult for conventional. Moreover, in application engineering, long-term forecasting and hydrologic model calibration, keys of area of hydrologic Forcasting, have been focuses of hydrology for a long time. The main research works are listed as following:1. The drawback of indirect multi-step-ahead prediction is error cumulation. In order to tackle this problem, a three-stage model, which is based on interpolation, time-delay technology, and adaptive time-delay neural network (ATNN), is presented in the paper. With interpolation algorithm, some dynamic virtual data are inserted into the original sequences at the point far from the current spot. So the impact of last prediction errors that would be iterated into the model for the next step prediction is decreased. And then the reliability of this indirect multi-step-ahead prediction model is obtained. Moreover, the three-stage model is specially fit for short specimen sequences. The tests on two benchmark nonlinear time series show the efficiency of our model.2. An EMD_DRNN model, which is based on empirical mode decomposition (EMD) and chaos analysis, is presented for direct multi-step-ahead prediction. First, we employed EMD to analysis original sequences into many basic modal partitions which can significantly represent potential information of original time series. For each partition, the algorithm is able to decrease the nonlinear complexity within each sequence, and cut the difference between each step prediction. By the process, the model obtains the capacity to learn various objective functions. Next, chaos features of those data sequences can be used to design DRNN, the main prediction parts. Finally, combination prediction can be performed and tested by some benchmark time series. Tested on time series of benchmark time series (yearly sunspotss and Mackey-Glass time series), models can provide more accurate result for multi-step prediction.3. In order to deal with multi-peak-value function optimization, an improved information-entropy-based immune algorithm (IIIA) for multimodal optimization is presented, which employs improved entropy-based concentration selecting operator and...