Pervasive installation of smart meters opens new possibilities for advanced analytics of electricity consumption at the level of the individual household. One of the important tasks in various Smart Grid applications, from demand-response to emergency management, is the short-term electricity load forecasting at different scales, from an individual customer to a whole portfolio of customers. In this work we perform a quantitative evaluation of different machine learning methods for short-term (1 hour ahead and 24 hour ahead) electricity load forecasting at the individual and aggregate level. We discuss the relevant features that best help to improve forecasting accuracy, as well as the effectiveness of exploiting correlations between different customers. Furthermore, for aggregate forecast, we explore different clustering techniques that can be used to segment the whole customer portfolio and show that forecasting each cluster separately and then aggregating the forecast produces better accuracy compared to the traditional approach (which forecast directly the aggregate load). We also found that the improvement provided by this strategy changes as a function of total customers.