A novel XGBoost-based featurization approach to forecast renewable energy consumption with deep learning models

For energy suppliers, forecasting the energy demand with accuracy is essential. The current studies in the literature have employed various statistical and machine/deep learning forecasting methods to predict energy consumption. Although deep learning methods have been successfully applied in this context, their performance can be improved by incorporating statistical features representing the characteristics of time series. This study proposes a novel two-stage forecasting framework composed of data preprocessing and model building. The data preprocessing component extracts statistical features from the input data, and then an XGBoost regressor is utilized to obtain the importance of each feature. The model-building component uses the obtained features and the original input data to construct the forecasting model. We implement three forecasting models based on the proposed approach using two state-of-the-art deep learning models, including the temporal convolution neural network and Multi-head Attention. We empirically evaluate the proposed approach on two renewable energy consumption datasets. The results of experiments indicate that incorporating features is beneficial for temporal convolution neural network-based and Multi-head Attention-based deep learning models performance. This study significantly contributes to the existing models in the literature, as the combined methods improve on their regular variants and the benchmark models.