Short term wind prediction with an ARIMA model

Wind forecasting has traditionally relied on computationally intensive Numerical Weather Prediction (NWP) models, whose accuracy depends on choice of parameterization schemes, model resolution, input data quality, and numerical solution techniques. Recently, data-driven methods such as Auto-Regressive Integrated Moving Average (ARIMA), Artificial Neural Networks (ANN), and Support Vector Machines (SVM) have emerged as alternatives, particularly for wind speed forecasting. However, accurate prediction of wind direction remains equally critical for many operational applications. This study presents an ARIMA-based model for forecasting hourly wind speed and direction with a lead time of up to three days. Unlike conventional machine learning approaches, the model explicitly incorporates the effects of atmospheric turbulence. Validation using a two-year observational dataset at a single site shows that, at a one-day lead time, the mean absolute error (MAE) in wind direction is ≤ 60° for ~ 84% of the year, while at a three-days lead time, the same accuracy is achieved for ~ 75% of the year. The results demonstrate predictive skill comparable to NWP models but at a fraction of computational cost, highlighting the model’s suitability for applications in wind farm operations, aviation, and nuclear energy sector emergency preparedness.