Wind forecasting by quantile regression Dr Geoffrey Pritchard
Wind forecasting by quantile regression Dr. Geoffrey Pritchard University of Auckland
Short term (within 2 hours) • The persistence forecast (“no change”) is hard to beat by much. • Important to indicate situation-dependent uncertainty – awareness of risks – probabilistic forecast: scenarios, full pdf
Met. Service e. PD - the kitchen-sink approach
Persistence forecast: 30 min x 5 -2 hr TP-5 0 TP-4 TP-3 • Half-hourly data • Wind forecast is actual output in TP-5 TP-2 TP-1 +30 min TP • Actual wind observed
72149 observations (~4 years)
72149 observations (~4 years)
Quantile regression Model for the t-quantile (0 < t < 1) of the conditional distribution of a response variable: Q(t) = coefficients S b i (t) xi explanatory variables Our xi will be (nonlinear) spline basis functions of current power.
Quantile regression fitting • To fit observations yi : minimize S rt(yi where rt is the function t-1 t • Reduces to linear programming. - b i (t) xi )
Additional predictor variables • Improve (? ) quantile models by adding terms with – – – wind direction barometric pressure time of day recent power variability etc. • In single-scenario forecasting, get little improvement on persistence. • In scenario generation, extra variables may help identify low- and high-uncertainty situations.
91 34 34
- Slides: 21