Low aggregations of electric load profiles are more fluctuating, relative forecast errors are comparatively high, and it has been shown that different forecast models and feature configurations may be best suitable for specific households or buildings. However, at low aggregations, the monetary incentive for manual feature engineering and model selection is low, as benefits from forecast improvements are small. Convolutional Neural Networks (CNN) have proven to achieve high accuracy in an end-to-end fashion with minimal effort for manual feature selection. WaveNet, a CNN-based approach, has been developed to handle noisy time-series data for speech recognition and synthesis. In this work we explore if WaveNet is suitable for short-term forecasts of lowly aggregated electric loads. We find that WaveNet performs similarly to, and slightly better than, typical benchmark models for individual households, at the cost of higher model complexity. Preliminary experiments show that transfer learning can further improve results and decrease training times for individual households, as a pattern such as the correlation between outside temperature and load can be learned as general features. For aggregations of 10-200 households WaveNet improves most over the benchmarks, e.g., 13% compared to vanilla Artificial Neural Networks at 200 households, making it possibly suitable for aggregated load forecasting.