國立虎尾科技大學 |

Multi-step Prediction With Neural Networks.

紀錄類型:	書目-語言資料,手稿 : Monograph/item
正題名/作者:	Multi-step Prediction With Neural Networks./
作者:	Kelley, Joseph.
面頁冊數:	1 online resource (146 pages)
附註:	Source: Dissertations Abstracts International, Volume: 86-01, Section: B.
Contained By:	Dissertations Abstracts International86-01B.
標題:	Electrical engineering. -
電子資源:	click for full text (PQDT)
ISBN:	9798383206300

Multi-step Prediction With Neural Networks.
Kelley, Joseph.

Multi-step Prediction With Neural Networks. - 1 online resource (146 pages)

Source: Dissertations Abstracts International, Volume: 86-01, Section: B.

Thesis (Ph.D.)--Oklahoma State University, 2024.

Includes bibliographical references

This paper demonstrates methods for modeling nonlinear dynamic systems to perform multi-step prediction. These methods are suitable for the design of model predictive controllers (MPC), forecasting power plant loads, and forecasting products sold in stores. There are numerous techniques for modeling nonlinear system, but in the paper we concentrated on the iterative method [1] [2] [3] for multi-step prediction, which we refer to as the TMS-Iterative method. In the iterative method a single model is trained to minimize one-step prediction errors, then the trained model is used for multi-step prediction. But for the TMS-Iterative method a single model is trained to minimize the multi-step prediction errors. In [4] and [5] it was suggested the TMS-Iterative method produces more robust long-term predictions. We found this to be especially true when modeling nonlinear dynamic systems.The paper is comprised of 4 sections. Each section presents a different recurrent neural network (RNN) architecture to perform multi-step prediction using the TMS-Iterative method. The 4 sections are (1) nonlinear input-output models, (2) nonlinear state-space models, (3) sequence-to-sequence (Seq2Seq) models, and (4) seasonal nonlinear model. In the first section we provide an extensive comparison of Nonlinear AutoRegressive model with eXogenous inputs (NARX) and Nonlinear AutoRegressive Moving Average model with eXogenous inputs (NARMAX) for multi-step predictions. We discuss in detail how the NARMAX models provide significantly improved forecasts for shorter prediction horizons, while the predictions converge as the horizon increases. In the second section we introduce the Nonlinear Innovation Form (NIF) and Nonlinear Output Form (NOF), which are nonlinear state-space models. Both the NIF and NOF are based on the Kalman filter, which estimates the hidden states in an RNN model. By estimating the hidden states we can improve the short-term and long-term forecasting accuracy. In the third section we demonstrate Seq2Seq models are fixed-weight adaptive and can learn time-varying dynamics. Seq2Seq models were originally introduced for national language processing. In the paper we extend the Seq2Seq to the modeling of time-varying nonlinear systems. In the fourth section the Periodic Nonlinear AutoRegressive model (PNAR) is introduced. The PNAR model is trained to learn seasonal variations in nonlinear dynamic systems. Seasonal variations are patterns that occur at consistent time intervals in a predictable fashion. In each section, the RNN models are compared using simulated and experimental data.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798383206300Subjects--Topical Terms:

596380
Electrical engineering.
Subjects--Index Terms:

Multi-step predictionIndex Terms--Genre/Form:

554714
Electronic books.

Multi-step Prediction With Neural Networks.
LDR:04038ntm a22004337 4500 001 1148367
005 20240924101907.5
006 m o d
007 cr bn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798383206300
035 $a (MiAaPQ)AAI31142080
035 $a AAI31142080
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Kelley, Joseph. $3 1474315
245 1 0 $a Multi-step Prediction With Neural Networks.
264 0 $c 2024
300 $a 1 online resource (146 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-01, Section: B.
500 $a Advisor: Hagan, Martin.
502 $a Thesis (Ph.D.)--Oklahoma State University, 2024.
504 $a Includes bibliographical references
520 $a This paper demonstrates methods for modeling nonlinear dynamic systems to perform multi-step prediction. These methods are suitable for the design of model predictive controllers (MPC), forecasting power plant loads, and forecasting products sold in stores. There are numerous techniques for modeling nonlinear system, but in the paper we concentrated on the iterative method [1] [2] [3] for multi-step prediction, which we refer to as the TMS-Iterative method. In the iterative method a single model is trained to minimize one-step prediction errors, then the trained model is used for multi-step prediction. But for the TMS-Iterative method a single model is trained to minimize the multi-step prediction errors. In [4] and [5] it was suggested the TMS-Iterative method produces more robust long-term predictions. We found this to be especially true when modeling nonlinear dynamic systems.The paper is comprised of 4 sections. Each section presents a different recurrent neural network (RNN) architecture to perform multi-step prediction using the TMS-Iterative method. The 4 sections are (1) nonlinear input-output models, (2) nonlinear state-space models, (3) sequence-to-sequence (Seq2Seq) models, and (4) seasonal nonlinear model. In the first section we provide an extensive comparison of Nonlinear AutoRegressive model with eXogenous inputs (NARX) and Nonlinear AutoRegressive Moving Average model with eXogenous inputs (NARMAX) for multi-step predictions. We discuss in detail how the NARMAX models provide significantly improved forecasts for shorter prediction horizons, while the predictions converge as the horizon increases. In the second section we introduce the Nonlinear Innovation Form (NIF) and Nonlinear Output Form (NOF), which are nonlinear state-space models. Both the NIF and NOF are based on the Kalman filter, which estimates the hidden states in an RNN model. By estimating the hidden states we can improve the short-term and long-term forecasting accuracy. In the third section we demonstrate Seq2Seq models are fixed-weight adaptive and can learn time-varying dynamics. Seq2Seq models were originally introduced for national language processing. In the paper we extend the Seq2Seq to the modeling of time-varying nonlinear systems. In the fourth section the Periodic Nonlinear AutoRegressive model (PNAR) is introduced. The PNAR model is trained to learn seasonal variations in nonlinear dynamic systems. Seasonal variations are patterns that occur at consistent time intervals in a predictable fashion. In each section, the RNN models are compared using simulated and experimental data.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Electrical engineering. $3 596380
650 4 $a Computer engineering. $3 569006
650 4 $a Computer science. $3 573171
650 4 $a Systems science. $3 1148479
650 4 $a Information technology. $3 559429
653 $a Multi-step prediction
653 $a NARMAX models
653 $a NARX
653 $a Recurrent neural network
653 $a Seq2Seq models
653 $a Time-series modeling
655 7 $a Electronic books. $2 local $3 554714
690 $a 0544
690 $a 0489
690 $a 0984
690 $a 0464
690 $a 0790
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a Oklahoma State University. $b Electrical Engineering. $3 1182179
773 0 $t Dissertations Abstracts International $g 86-01B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31142080 $z click for full text (PQDT)