On the mixture of wind speed distribution in a Nordic region

The assessment of wind energy potential at sites of interest requires reliable estimates of the statistical characteristics of wind speed. A probability density function (pdf) is usually fitted to short-term observed local wind speed data. It is common for wind speed data to present bimodal distributions for which conventional one-component pdfs are not appropriate. Mixture distributions represent an appropriate alternative to model such wind speed data. Homogeneous mixture distributions remain rarely used in the field of wind energy assessment while heterogeneous mixture models have only been developed recently. The present work aims to investigate the potential of homogeneous and heterogeneous mixture distributions to model wind speed data in a northern environment. A total of ten two-component mixture models including mixtures of gamma, Weibull, Gumbel and truncated normal are evaluated in the present study. The estimation of the parameter of the mixture models are obtained with the least-squares (LS) and the maximum likelihood (ML) methods. The optimization of the objective functions related to these estimation methods is carried out with a genetic algorithm that is more adapted to mixture distributions. The case study of the province of Québec (Canada), a Northern region with an enormous potential for wind energy production, is investigated in the present work. A total of 83 stations with long data records and providing a good coverage of the territory of the province are selected. To identify the most appropriate one-component distribution for the selected stations, the newly proposed method of L-moment ratio diagram (MRD) is used. The advantages of this approach are that it is simple to apply and it allows an easy comparison of the fit of several pdfs for several stations on a single diagram. One-component distributions are compared with the selected mixture distributions based on model selection criteria. Results show that mixture distributions often provide better fit than conventional one-component distributions for the study area. It was also observed that the ML method outperforms the LS method and that the mixture model combining two Gumbel distributions using ML is the overall best model.