Experimental seasonal forecasts of low summer flows in the River Thames, UK, using Expert Systems

摘要

ABSTRACT: Empirical seasonal forecasts of hydroclimatic variables are typically limited by unrealistic assumptions about stationarity, linearity and normality. Although empirical, Expert Systems (ES) are not bound by such formal data constraints and may be trained on continuous and categorical data simultaneously. Using an ES, summer flows for the River Thames, UK (1971 to 2001), were forecast with up to 6 mo lead-time from climate data from the preceding winter months. The ES models successfully forecast the correct sign of August flow anomalies up to 77% of the time and June to August anomalies 71% of the time. Forecasts of low flow (lowest quartile of the long-term average) in August and summer as a whole were correctly forecast 79 and 89% of the time, respectively. The models also correctly predicted below-average flows for the drought years of 1976, 1984, 1989 to 1992, 1995 and 1997. These extreme events would not have been captured by a forecast based on long-term climatology alone.