Non-stationary statistical modelling of wind speed: A case study in eastern Canada

Highlights: • Classical modeling approaches do not take into account interannual variability and trends. • The proposed approach provides wind speed distribution conditionally on a set of predictors. • Annual goodness-of-fit at the studied stations improved on average with the non-stationary model. • Influential climatic indices are used as predictors in the non-stationary model. The assessment of wind energy potential is generally based on the analysis of the statistical distribution of observed wind speed of short time resolution. Record periods of observational data used in practice at sites of interest are often very short, often ranging from a few months to a few years. Predictions based on such small record periods are likely to be biased as it is recognized that wind speed is subject to important interannual variability and long-term trends. Large-scale atmospheric circulation patterns have an important influence on wind speed. Their predictable nature can make them useful for the prediction of wind speed during the lifetime of wind farm projects. This feature is not exploited in practice. It is proposed in this study to introduce predictors of the wind speed in non-stationary statistical models. This approach allows the development of predictions of the wind speed distribution conditionally on the state of the predictors. The predictors used here are indices of atmospheric circulation to account for the interannual variability and a temporal index to account for the long-term temporal trend. The proposed approach was applied to hourly wind speed data at selected meteorological stations in the province of Québec (Canada). 20 stations with long record periods of over 30 years of data were used. The most important circulation indices identified in the study area are the North-Atlantic Oscillation (NAO) during the winter season and the Pacific North American (PNA) during the spring season. Results indicate that the annual goodness-of-fit at the stations of the case study improved on average when the non-stationary model is used compared to the stationary model. The proposed approach can potentially be used to model wind speed during the projected lifetime of wind farms using forecasts of the predictors.