This study investigates the impact of various attributes on forecasting daily electricity consumption within Romania’s national grid. Time-series data comprising instantaneous power values, recorded by the National Grid Operator, were extracted from their continuously updated database for the period spanning 2013 to 2021. A set of exogenous variables—including meteorological parameters represented by daily averaged weather conditions, calendar-related data, and an energy policy measure (Daylight-Saving Time)—was integrated. The resulting dataset was used to train five deep learning architectures, both univariate and multivariate. Explainability techniques SHAP and LIME were employed to derive local and global interpretability insights. Both methods indicated the day type as the most influential factor. Attributes traditionally regarded as critical in shaping electricity load profiles, such as temperature, were ranked lower in terms of influence. Additionally, it was observed that LIME is more effective for generating localized instance-specific explanations, whereas its suitability for global interpretability remains limited.