Wind speed forecasting using univariate and multivariate time series models

Accurate short-term wind speed forecasts are essential for optimizing wind energy harvesting and maintaining grid reliability. This study evaluates the SARIMA, SARIMAX, VAR, and VARMA time series models, using hourly and sub-hourly wind speed and direction data from four Algerian sites with Mediterranean and Saharan climates. The forecasting process incorporated both simultaneous and rolling forecasts approaches, with LSTM serving as a benchmark. Simultaneous forecasts showed that LSTM outperformed the statistical models, by at least 46.95% in RMSE and 48.22% in MAE. However, the rolling forecast approach significantly improved the statistical models' performance, mitigating the plateau effect common in time series predictions, despite being computationally expensive. Notably, SARIMA emerged as the most reliable model overall, while VARMA offering a competitive alternative, whereas VAR often lagged behind. Despite LSTM's initial dominance, it was ultimately surpassed by these models by at least of 65.54% and 58.31% for RMSE and MAE respectively, highlighting the effectiveness of rolling forecasts approach. Moreover, we observed that incorporating the sine and cosine of hours alongside the wind direction as exogenous variables enhanced the performance of SARIMAX. Furthermore, VAR and VARMA excelled in locations exhibited significant correlation between wind features. Notably, The largest RMSE and MAE values occurred in areas with low data quality and high volatility.