NON-STATIONARY ANALYSIS OF ANNUAL MAXIMUM STREAMFLOW OF BAITARANI RIVER BASIN

Abstract

As many researchers arouse the concern for the elements that can affect the frequency and magnitude of the extreme, the concept of non-stationarity came into the picture. The typical analysis (generally flood frequency analysis) assumes the data to be independent and stationary. Putting in other words, the stationary analysis assumes the statistical parameters to be time-invariant. Stationary analysis still overpowers the non-stationary analysis in some cases, as the latter involves many parameters adding complexity to the models, and it is still an argument about adopting among the both. So, underestimating or overestimating may cause the associated design flood to affect the functioning of the hydraulic structures, infrastructures, water management and public safety. Evidence in some literature indicates the variables, namely, climate change, anthropogenic activities impacting the magnitude and frequency of flood. To analyse in the context of variables, the non-stationary analysis provides a better understanding of the parameters which are likely to vary with respect to the variables. The variables can be time, climate indices; urbanisation which has been carried out in many literature as covariates to the response variable. In this case study time has been considered as the covariate for the non-stationary analysis, ignoring the presence of any other covariates. For this work, three gauge stations, namely, Anandpur, Champua and Akhuapada of the Baitarani river basin, have been considered. Preliminary analysis like time series analysis was done to have an understanding of adopting the models between stationary or non-stationary. Non-parametric tests like Mann-Kendall trend, Cox and Stuart Trend, etc. were used for trend detection. Similarly, for change point detection and randomness, other non-parametric tests were performed on stations of the Baitarani basin. Due to the uncertain behaviour of the time-series like randomness, it was preferred to take up non-stationary analysis. The statistical frameworks adopted were GAMLSS (Generalised additive models for location, scale and shape parameters) and Pro-NEVA (Process-informed non-stationary analysis) for non-stationary analysis. GAMLSS tool provides a large range of distribution, making it suitable for the analysis, while Pro-NEVA is built upon the hybrid evolution Markov Chain Monte Carlo approach. In GAMLSS approach, continuous two-parameter distribution like Gamma, Gumbel, Inverse Gaussian, Normal, Reverse Gumbel, Weibull, Lognormal, are adopted, which are widely used in the field of hydrology. Lognormal and Gamma distribution were found to be best fitted to the observation. GAMLSS linear models were adopted based on the goodness of fit criteria, which indicates to adopt stationary models in the Baitarani basin. In Pro-NEVA, GEV and LP3 models were adopted for the analysis.