Online Regime Switching Vector Autoregression Incorporating Spatio-temporal Aspects for Short Term Wind Power Forecasting

Abstract: This master thesis examines short term wind power forecasting time series models focusing on regimes conditioned to meteorological conditions and the incorporation of spatio-temporal aspects. Novel regime switching autoregressive and vector autoregressive models are proposed, implemented in a .NET framework, and evaluated. The vector autoregressive framework takes advantage of cross-correlation between sites incorporating upstream online production information from all wind farms within a given region. The regimes are formed using K-means clustering based on forecast meteorological conditions. Each of the proposed models are fit to hourly historical data from all of 2015 for 24 wind farms located in Sweden and Finland. Forecasts are generated for all of 2016 and are evaluated against historical data from that year for each of the 24 wind farms. The proposed models are successfully implemented into the .NET framework of Vitec Software’s Aiolos Forecast Studio, which is widely used in the Northern and Western Europe. Vitec’s Aiolos wind power forecast model is based on an ensemble of numerical weather prediction models and adaptive statistical machine learning algorithms. The proposed models are found to have significantly lower mean absolute error and root mean squared error compared to the Aiolos model and autoregressive model benchmarks. The improved short term wind power forecast will inform operation and trading decisions and translate to significant reductions in balancing costs for Vitecs customers. The improvement is valued at as much as between 9.4 million Euros to 42.3 million Euros in reduced balancing costs. Spatio-temporal aspects for wind power forecasting is shown to continue to be promising for improving current state-of-the-art wind power forecasting.