Short-term electricity consumption forecasting plays a crucial role in the optimal operation of energy systems and smart grids. Despite the widespread success of machine learning models in improving forecasting accuracy, their lack of interpretability often hinders trust and application in energy decision-making. This paper presents an interpretable framework for hourly electricity consumption prediction using lagged load and temperature features. To this end, the performance of linear and non-linear models is systematically compared on a one-year hourly electricity consumption dataset, including ambient temperature and workday information for a commercial building. To enhance model transparency, Explainable Artificial Intelligence methods are employed to analyze feature importance at both global and local levels. Results indicate that short-term and daily lags of electricity consumption play a dominant role in consumption dynamics, while the effect of temperature is non-linear and often appears with a time delay. Tree-based models significantly outperform the linear model and demonstrate a higher ability to identify complex interactions and lagged effects of temperature. SHapley Additive exPlanations analysis reveals consistent patterns in feature importance, and its waterfall plots and Local Interpretable Model Agnostic Explanations provide complementary local insights under different consumption conditions. The proposed framework, while improving forecasting accuracy, provides a transparent and meaningful interpretation of electricity consumption behavior and can be used as a reliable tool in energy planning and demand-side management.