Journal of Water and Climate Change (Dec 2022)
Performance evaluation of univariate time-series techniques for forecasting monthly rainfall data
Abstract
In this article, the performance evaluation of four univariate time-series forecasting techniques, namely Hyndman Khandakar-Seasonal Autoregressive Integrated Moving Average (HK-SARIMA), Non-Stationary Thomas-Fiering (NSTF), Yeo-Johnson Transformed Non-Stationary Thomas-Fiering (YJNSTF) and Seasonal Naïve (SN) method, is carried out. The techniques are applied to forecast the rainfall time series of the stations located in Kerala. It enables an assessment of the significant difference in the rainfall characteristics at various locations that influence the relative forecasting accuracies of the models. Along with this, the effectiveness of Yeo-Johnson transformation (YJT) in improving the forecast accuracy of the models is assessed. Rainfall time series of 18 stations in Kerala, India, starting from 1981 and ending in 2013, is used. A classification system based on root mean square error (RMSE), mean absolute error (MAE) and Nash–Sutcliffe model efficiency coefficient (NSE) is proposed and applied to find the best forecasting model. The models HK-SARIMA and YJNSTF performed well in the Western lowlands and Eastern highlands. In the Central midlands, out of 12 stations, the performance indices of 8 stations are in favour of the HK-SARIMA model. It can be concluded that HK-SARIMA models are more reliable for forecasting the monthly rainfall of the stations located in all geographic regions in the state of Kerala. HIGHLIGHTS YJT is applied to transform the non-normal rainfall time series more Gaussian-like distribution and simultaneously increases the TF model's forecasting ability.; The HK algorithm used in this study connects the unit root test, minimising the corrected Akaike information criterion (AICc) and maximum likelihood estimator (MLE) to obtain the model order and parameter (coefficients) of a SARIMA model.; The hyper-parameters of SARIMA models obtained from the HK algorithm identified that the seasonal autoregressive (AR) and moving average (MA) components are more prominent than the non-seasonal components.; The concept of strength of seasonality and the strength of trend is utilised to explore the non-stationary nature of rainfall datasets, and it was inferred that all eighteen stations are indeed non-stationary.; HK-SARIMA performs better than YJNSTF, NSTF and SN, which can be attributed to the fact that HK-SARIMA handles seasonality better than TF by creating an auto-regression equation on the time series dataset.;
Keywords
