Wind Forecasting In Railway Engineering PDF Download

Short-term wind and wind power forecasts are required for the reliable and economic operation of power systems with significant wind power penetration. This thesis presents new statistical techniques for producing forecasts at multiple locations using spatiotemporal information. Forecast horizons of up to 6 hours are considered for which statistical methods outperform physical models in general. Several methods for producing hourly wind speed and direction forecasts from 1 to 6 hours ahead are presented in addition to a method for producing five-minute-ahead probabilistic wind power forecasts. The former have applications in areas such as energy trading and defining reserve requirements, and the latter in power system balancing and wind farm control. Spatio-temporal information is captured by vector autoregressive (VAR) models that incorporate wind direction by modelling the wind time series using complex numbers. In a further development, the VAR coefficients are replaced with coefficient functions in order to capture the dependence of the predictor on external variables, such as the time of year or wind direction. The complex-valued approach is found to produce accurate speed predictions, and the conditional predictors offer improved performance with little additional computational cost. Two non-linear algorithms have been developed for wind forecasting. In the first, the predictor is derived from an ensemble of particle swarm optimised candidate solutions. This approach is low cost and requires very little training data but fails to capitalise on spatial information. The second approach uses kernelised forms of popular linear algorithms which are shown to produce more accurate forecasts than their linear equivalents for multi-step-ahead prediction. Finally, very-short-term wind power forecasting is considered. Five-minute-ahead parametric probabilistic forecasts are produced by modelling the predictive distribution as logit-normal and forecasting its parameters using a sparse-VAR (sVAR) approach. Development of the sVAR is motivated by the desire to produce forecasts on a large spatial scale, i.e. hundreds of locations, which is critical during periods of high instantaneous wind penetration.