Stochastic model stationarization by eliminating the periodic term and its effect on time series prediction
| Autor: | Seyed Ehsan Fatemi, Hamid Moeeni, Hossein Bonakdari |
|---|---|
| Rok vydání: | 2017 |
| Předmět: |
010504 meteorology & atmospheric sciences
Series (mathematics) Stochastic modelling 0208 environmental biotechnology 02 engineering and technology 01 natural sciences 020801 environmental engineering Term (time) Amplitude Streamflow Climatology Spectral analysis Autoregressive integrated moving average Time series 0105 earth and related environmental sciences Water Science and Technology Mathematics |
| Zdroj: | Journal of Hydrology. 547:348-364 |
| ISSN: | 0022-1694 |
| DOI: | 10.1016/j.jhydrol.2017.02.012 |
| Popis: | Because time series stationarization has a key role in stochastic modeling results, three methods are analyzed in this study. The methods are seasonal differencing, seasonal standardization and spectral analysis to eliminate the periodic effect on time series stationarity. First, six time series including 4 streamflow series and 2 water temperature series are stationarized. The stochastic term for these series obtained with ARIMA is subsequently modeled. For the analysis, 9228 models are introduced. It is observed that seasonal standardization and spectral analysis eliminate the periodic term completely, while seasonal differencing maintains seasonal correlation structures. The obtained results indicate that all three methods present acceptable performance overall. However, model accuracy in monthly streamflow prediction is higher with seasonal differencing than with the other two methods. Another advantage of seasonal differencing over the other methods is that the monthly streamflow is never estimated as negative. Standardization is the best method for predicting monthly water temperature although it is quite similar to seasonal differencing, while spectral analysis performed the weakest in all cases. It is concluded that for each monthly seasonal series, seasonal differencing is the best stationarization method in terms of periodic effect elimination. Moreover, the monthly water temperature is predicted with more accuracy than monthly streamflow. The criteria of the average stochastic term divided by the amplitude of the periodic term obtained for monthly streamflow and monthly water temperature were 0.19 and 0.30, 0.21 and 0.13, and 0.07 and 0.04 respectively. As a result, the periodic term is more dominant than the stochastic term for water temperature in the monthly water temperature series compared to streamflow series. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|
Full Text from ScienceDirect