Deterministic Wind Speed Forecasting using Gradient Boosting Models

The wind is highly intermittent in nature and obtaining accurate wind forecasts is the need of the electricity markets for decision making, planning and maintenance activities. In the era of machine learning, an intelligent model is required to learn non-linear interactions of the wind speed with less complexity and high efficiency. Gradient boosting machines (GBM) produce reliable predictions without much pre-processing and are natural in various forecasting competitions. In this paper, we compare three prevalent GB based ensemble methods for wind speed forecasting, which are Extreme Gradient Boosting (XGBoost), Light Gradient Boosting (LGB), and Categorical Gradient boosting (CatBoost) machine models. In XGBoost, regularisation of the objective function is done with its capability of handling missing values and sparsity, due to which this method shows high accuracy. In LGB, weights are assigned to the smaller and larger gradients in descending order, and then only larger gradients are considered, which makes it very fast in computation. In CatBoost, training data is selected by random permutation of the entire data-set, which makes it faster when compared with XGBoost. These methods are tested on the GEFCOM-14 wind speed dataset and are compared with other benchmark methods.