While there is an increasing interest in Machine Learning (ML) based solutions, scarce research has been devoted to the deployment and monitoring of ML models. In this work, we address this research gap by proposing a new data drift ML update strategy that only considers changes in the input features. Using the realistic Growing Window (GW) and Rolling Window (RW) ML deployment simulation schemes, we propose two Drift variants (DGW and DRW), which are compared with three other ML update approaches: Single Training (ST) and Periodic retraining methods (PGW and PRW). Several computational experiments were held, using the XGBoost regression learner and 8 public-domain datasets related to energy production and consumption. Overall, when considering both the predictive performance and computational effort, the proposed DGW and DRW obtained competitive results. In particular, quality predictive errors were achieved (overall value of 1.33% for DGW and 1.43% for DRW), while requiring around half of the computational effort when compared with the periodic update versions (PGW and PRW).