Modified Relative Strength Effect Analysis In Neural Network Based Hydrologic Modeling

Ningyu Song, Purdue University

Abstract

In the last decade, many artificial neural network (ANN) based models were developed and used in different hydrological modeling. Trained ANN model has a highly non linear equation. Hence, in the past, while using neural network for functional approximation, most of the researches were conducted by considering neural network as a black box model. Further, while deciding the architecture of the neural network model as well as for identifying essential inputs, trial and error based approach was adopted systematically. This study uses modified relative strength effect (RSE) in deciding essential inputs and provides more insight about the usefulness of ANN model. Efficacy of RSE as partial autocorrelation coefficient in time series modeling is also examined in this study. One of the most important components of neural network modeling is prescribing an appropriate training termination criterion. This step avoids over training of the neural network model. The traditional termination criterion recommends for splitting the dataset into training, testing, and validation sets. Yet for many real world applications, this way is expensive and luxury for the database people finally got. Frequently small datasets were encountered, which are very expensive to generate. This study proposes a new termination criterion approach using which the whole dataset will be split into training and validation only, without splitting the dataset into three parts. This procedure is based on RSE and Relative Strength Index (RSI). Revised approach used for finding RSE is very suitable for automating the training termination. Proposed model was tested using four different datasets and the results are very promising. This research was also extended to Self Organizing Maps (SOM) to identify appropriate network topology. Suitability of SOM in regional trend analysis was examined using case study involving Indiana Reservoir inflows. This study indicated the usefulness of SOM in identifying regional trends.

Degree

M.S.E.

Advisors

Yang, Purdue University.

Subject Area

Computer Engineering|Artificial intelligence

COinS