ACCURACY OF FILLING STREAMFLOW GAPS USING ARTIFICIAL INTELLIGENCE

Abstract

GAPS IN STREAMFLOW DATA ARE COMMON EVEN THOUGH CONTINUOUS TIME-SERIES ARE REQUIRED FOR WATER BALANCE CALCULATIONS. THIS STUDY EXPLORES HOW DIFFERENT FACTORS AFFECT THE ACCURACY OF FILLING STREAMFLOW GAPS IN THE 75 SQUARE KILOMETER MAE SA CATCHMENT IN NORTHERN THAILAND. MODELS ARE BUILT WITH FREE SOFTWARE AND USE DIFFERENT COMBINATIONS OF AN UNUSUALLY RICH VARIETY OF DATA AS INPUT, INCLUDING SPATIAL RAINFALL AND SOIL MOISTURE, 3 HYDROLOGIC AND 7 CLIMATE VARIABLES.

ACCURACY LINEARLY DECREASES WITH GAP LENGTH FOR 1-50 DAY GAPS. SUPPORT VECTOR MACHINE IS MUCH MORE ACCURATE THAN THE OTHER TECHNIQUES FOR 1-50 DAY GAPS. FOR GAPS LONGER THAN 59-DAYS, THE SIMPLE CONCEPTUAL IHACRES MODEL IS MORE SUITABLE. A PROMISING NEW PROXIMAL TRAINING METHOD IS INTRODUCED THAT RESTRICTS MODEL TRAINING DATA TO N DAYS BEFORE AND AFTER EACH GAP. IT INCREASES AVERAGE ANN ACCURACY (MNSE) FROM 0.60 TO 0.67 (N=150). PROXIMAL TRAINING ALSO SUBSTANTIALLY REDUCES SUPPORT VECTOR MACHINE RUNNING TIME FROM 29MIN TO 1