Ship Model Identification with Genetic Algorithm Tuning
Modeling is the most important component in predictive controller design. It should predict outputs precisely and fast. Thus, it must be adequate for the ship dynamics while having as simple a structure as possible. In a good ship model the standard deviation of a particular coefficient should not e...
Gespeichert in:
Veröffentlicht in: | Applied sciences 2021-06, Vol.11 (12), p.5504 |
---|---|
1. Verfasser: | |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | Modeling is the most important component in predictive controller design. It should predict outputs precisely and fast. Thus, it must be adequate for the ship dynamics while having as simple a structure as possible. In a good ship model the standard deviation of a particular coefficient should not exceed 10% of its value. Fitting the validation data to 80% for short-term prediction and 65% for long-term prediction is treated as a declared benchmark for model usage in ship course predictive controller. Regularization was proposed to ensure better state-space models to fit the real ship dynamics and more accurate standard deviation value control. Usage of the simulation results and real-time trials, as model estimation and validation data, respectively, during the identification procedure is proposed. In the first step a predictive linear model is identified conventionally, and then coefficients are regularized, based on the validation data, using a genetic algorithm. Particular linearized model coefficient standard deviations were decreased from more than 100% of their values to approximately 5% of them using genetic algorithm tuning. Moreover, the proposed method eliminated model output signal oscillations, which were observed during the validation process based on experimental data, gained during ship trials. Improved mapping of ship dynamics was achieved. Fit to validation data increased from 71% and 54% to 89% and 76%, respectively, for short-term and long-term prediction. The proposed method, which may be applied to real applications, is easily applicable and reliable. The tuned model is sufficiently suited to plant dynamics and may be used for future predictive control purposes. |
---|---|
ISSN: | 2076-3417 2076-3417 |
DOI: | 10.3390/app11125504 |