A comprehensive benchmark of machine learning-based algorithms for medium-term electric vehicle charging demand prediction

Abstract The current difficulties faced by evolutionary smart grids, as well as the widespread electric vehicles (EVs) into the modernised electric power system, highlight the crucial balance between electricity generation and consumption. Focusing on renewable energy sources instead of fossil fuels can provide an enduring environment for future generations by mitigating the impacts of global warming. At this time, the popularity of EVs has been ascending day by day due to the fact that they have several advantages such as being environmentally friendly and having better mileage performance in city driving over conventional vehicles. Despite the merits of the EVs, there are also a few disadvantages consisting of the integration of the EVs into the existing infrastructure and their expensiveness by means of initial investment cost. In addition to those, machine learning (ML)-based techniques are usually employed in the EVs for battery management systems, drive performance, and passenger safety. This paper aims to implement an EV monthly charging demand prediction by using a novel technique based on an ensemble of Pearson correlation (PC) and analysis of variance (ANOVA) along with statistical and ML-based algorithms including seasonal auto-regressive integrated moving average with exogenous variables (SARIMAX), convolutional neural networks (CNNs), extreme gradient boosting (XGBoost) decision trees, gated recurrent unit (GRU) networks, long short-term memory (LSTM) networks, bidirectional LSTM (Bi-LSTM) and GRU (Bi-GRU) networks for the Eastern Mediterranean Region of Türkiye. The performance and error metrics, including determination coefficient ( R $$^2$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mmultiscripts> <mml:mrow/> <mml:mrow/> <mml:mn>2</mml:mn> </mml:mmultiscripts> </mml:math> ), mean absolute percentage error (MAPE), mean absolute error (MAE), and mean absolute scaled error (MASE), are evaluated in a benchmarking manner. According to the obtained results, in Scenario 1, a hybrid of PC and XGBoost decision trees model achieved an R $$^2$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mmultiscripts> <mml:mrow/> <mml:mrow/> <mml:mn>2</mml:mn> </mml:mmultiscripts> </mml:math> of 96.21%, MAPE of 5.52%, MAE of 6.5, and MASE of 0.195 with a training time of 2.08 s and a testing time of 0.016 s. In Scenario 2, a combination of ANOVA and XGBoost decision trees model demonstrated an R $$^2$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mmultiscripts> <mml:mrow/> <mml:mrow/> <mml:mn>2</mml:mn> </mml:mmultiscripts> </mml:math> of 96.83%, a MAPE of 5.29%, a MAE of 6.0, and a MASE of 0.180 with a training time of 1.62 s and a testing time of 0.012 s. These findings highlight the superior accuracy and computational efficiency of the XGBoost models for both scenarios compared to others and reveal XGBoost's suitability for EV charging demand prediction.