Estimation of prediction interval in ANN-based multi-GCMs downscaling of hydro-climatologic parameters

• Published in: Journal of hydrology (Amsterdam) Vol. 579; p. 124226
• Main Authors: Nourani, Vahid, Jabbarian Paknezhad, Nardin, Sharghi, Elnaz, Khosravi, Abbas
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2019
• Subjects: Artificial neural network; Downscaling; General circulation models; Prediction interval; Downscaling; Artificial neural network; General circulation models; Prediction interval
• ISSN: 0022-1694; 1879-2707
• DOI: 10.1016/j.jhydrol.2019.124226

•Lower Upper Bound Estimation method used for constructing prediction intervals.•General circulation models applied for downscaling precipitation and temperature.•Results compared with the results of classic Bootstrap method.•Performance of methods was assessed by the criteria of coverage and width.•Prediction intervals for temperature were narrower than precipitation. In this paper, point prediction and prediction intervals (PIs) of artificial neural network (ANN) based downscaling for mean monthly precipitation and temperature of two stations (Tabriz and Ardabil in North West of Iran) were evaluated using general circulation models (GCMs). PIs were constructed by novel Upper Lower Bound Estimation (LUBE) method in which an ANN with two outputs was constructed for estimating the prediction bounds. Also, Bootstrap method as a classic technique for assessing uncertainty of ANN was used to further examine the proposed LUBE method. In this way, the accuracy of PIs was quantified by coverage and width criteria. Three GCMs, Can-ESM2, BNU-ESM, INM-CM4 and ensemble-GCM (ensemble of mentioned models) were used in four grid points around each of station for evaluating ANN-based downscaling of precipitation and temperature parameters. Comparison between the results of two methods indicated that LUBE method could lead to more reliable results than the Bootstrap method. PIs width and coverage probability were 10–40% lower and 2–10% higher than the Bootstrap method for different GCMs, respectively. Ensemble-GCM led to more accurate results so that computed PIs width and coverage probability were 10–60% lower and 2–20% higher than those for the single GCMs.