Wind Speed Forecasting with Missing Values

Abstract

In this study, a new short term wind speed forecasting approach, which uses long term past observations, is proposed. The performance assessment of the wind speed forecasting framework is carried out using real data from meteorological stations in Marmara region of Turkey. The data sets are not complete due to equipment failures. The proposed approach builds on data de-trending, covariance-factorization via a subspace method, and one-step-ahead and multi-step-ahead Kalman filter prediction ideas. It is shown that trimming of diurnal, weekly, monthly, and annual patterns in data significantly enhances estimation accuracy. Experimental test results demonstrate that the proposed multi-step-ahead forecasting outperforms the benchmark values computed with the persistent forecasting models.