Enhancing electric vehicle charging efficiency at the aggregator level: A deep-weighted ensemble model for wholesale electricity price forecasting

The proliferation of electric vehicle (EV) adoption strains low-voltage distribution networks, particularly in aggregated charging scenarios, prompting utility companies to incentivize charging aggregators for optimizing load balancing within thermal limits. These aggregators utilize machine learning algorithms to understand electricity price signals and orchestrate the optimization of the EV charging process. However, conventional machine learning approaches fall short when dealing with the dynamic and volatile nature of electricity prices, emphasizing the necessity for advanced ensemble models. This paper introduces a novel Deep-Weighted Ensemble Model (DWEM) rooted in standard and stacked Long Short-Term Memory (LSTM) networks designed for wholesale electricity price forecasting, to manage the EV charging at the aggregator level. The ensemble development process involves developing an architecture that highlights the significance of the DWEM model in supporting aggregators for the charging optimization of EVs. The charging optimization problem of aggregated EVs is formulated, and the heuristic mechanism is systematically presented, evaluating various weight configurations, and selecting those characterized by the highest levels of accuracy to comprise the ensemble model. Moreover, we incorporated a standard deviation mechanism to evaluate the impact of the proposed DWEM on forecasting accuracy, mean squared error, and mean absolute error across various standard deviation levels. We leveraged a publicly available Houston electricity dataset and performed a detailed data engineering mechanism, accounting for data both with and without outliers. Subsequently, we applied the proposed DWEM to this dataset and conducting three types of comparative analysis: (a) evaluating model performance in terms of accuracy, mean square error, and mean absolute error; (b) assessing aggregator charging analysis focusing on charging load and cost; and (c) analyzing computational complexity and execution time. The simulation results demonstrated a improvement in accuracy and reduction in charging load and cost compared to state-of-the-art methods, while maintaining competitive computational complexity.