Wind energy forecasts with supervised machine learning techniques

United Nations set clean and affordable energy (7th) sustainable development goals. Among renewables, wind energy requires frequent observations. Due to limited finance in developing countries like Bangladesh, accurate and wide observations are one of the major challenges in operation and maintenance for energy harvests. Thus, this paper developed energy forecast models with supervised machine learning (ML) algorithms: support vector regressor (SVR) and random forest (RF) and compared with linear regression (LR) model. With global wind atlas (GWA) datasets (heights: 10 to 200m), training and testing performed with SVR, RF, and LR models. Model performances evaluated with Pearson correlation, coefficient of determinations, mean absolute percentage (MAPE), and root mean square errors (RMSE). During training and testing, ML estimations outperformed LR model as MAPE indicated ≤15% and ≥50% respectively for global and local-scale datasets. For an agricultural land with limited observations, ML predictions were performed for Bogura city with NASA-POWER dataset representing an MAPE of 4.84% (SVR) and 0.91% (RF). At Bogura on average, the final output showed annual energy production of 189 Kilowatt where probability for wind speeds about 70% (3 to 6 m/s at 50m). In conclusion, ML model accuracy is efficient for successful energy operations in Bangladesh and other developing countries with limited observations.