Partially recurrent neural networks for filtering and forecasting | |
| Author | Nguyen Thi Hoang Yen |
| Call Number | AIT Thesis no.IM-02-07 |
| Subject(s) | Neural networks (Computer science) Forecasting--Computer simulation |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science, School of Advanced Technologies |
| Publisher | Asian Institute of Technology |
| Series Statement | Thesis ; no. IM-02-07 |
| Abstract | The applicability of pattially recurrent neural networks (PRNN's) for filtering and forecasting has been tested in this study. The natural data series, namely discharge time series, in the Red River Basin in Viet Nam have been filtered and forecasted with the two selected PRNNs: the recurrent neural network (RNN) with only feedback from hidden nodes (called PRNN-1) and the Elman RNN (called PRNN-II). The online learning algoritlun proposed by Chairatanatrai, in which the Error Back Propagation (EBP), Error Self-Recurrent (ESR), and Recursive Least Squares (RLS) methods are employed has been modified to apply to the two selected PRNN' s. The obtained results indicate that the modified algoritluns can be used to efficiently train the PRNN-I and PRNN-II for both forecasting and filtering purposes. The network structure much affects the performance of network model. For the best network performance, an appropriate network structure is required. Since the issue is still under research, the trial-and-error method has been used in many applications so far. In this study, a procedure for defining the appropriate network structure based on the Bayesian Information Criterion (BIC) has been proposed instead. The proposed procedure has been successfully applied to determine the appropriate structure for the RNN's considered in the present study. In order to be able to compare the performance of the selected PRNN's with the fully recurrent neural network (FRNN), forecasting and filtering the flow time series in the Red River Basin with the FRNN have been carried out as well. The experimental results show that the PRNN-II model performs best in forecasting while the best performance in filtering goes with the FRNN model. The performance of the three models, the FRNN, the PRNN-I and the PRNN-II is, in fact, quite comparable in forecasting with the same mean efficiency index (EI) of 0.93 in the validation phase and only few percents difference in the training phase. With high value of EI of over 0.90 for most experimental cases, one may conclude that the three models can well forecast the stream flow in the Red River Basin. Good results of filtering the stream flow data in the Basin have also been obtained. High value of EI of over 0.90 has been obtained for most cases of using FRNN and PRNN-I for filtering. However, lower values of EI with its mean of 0.86 are found when the PRNN-Il model is applied. |
| Year | 2002 |
| Corresponding Series Added Entry | Asian Institute of Technology. Thesis ; no. IM-02-07 |
| Type | Thesis |
| School | School of Advanced Technologies (SAT) |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Information Management (IM) |
| Chairperson(s) | Huynh Ngoc Phien; |
| Examination Committee(s) | Sadananda, Ramakoti;Hoang Le Tien; |
| Degree | Thesis (M.Sc.) - Asian Institute of Technology, 2002 |