Maximizing the energy flexibility in residential districts by supporting demand response programs is a promising vision for enabling renewable energy integration, such as distributed solar energy systems, and stabilizing power grids. In particular, residential districts consisting of a cluster of buildings, represent an interesting source of flexibility, with their large needs for heating and cooling, which can be supplied by a central energy plant. In this context, demand response programs motivate end-users to modulate their normal energy usage behavior in response to electricity prices or through corresponding payment incentives. Despite these benefits, the accurate estimation of baselines when activating demand response events remains a major challenge. To this end, this study proposes a deep learning approach to provide accurate estimates of the demand baseline of more than 300 dwellings, over a 24-hour forecast horizon. In this approach, the forecasting model learns from a dataset of energy demand cases representing various combinations of dwelling characteristics, working schedules, and occupant behaviors, simulated in an urban-energy modeling tool. The deep learning model performance is tested and compared with traditional machine learning models, such as extreme gradient boosting and support vector regression. Subsequently, the trained models for this approach are used to predict future energy reductions over a 2-hour time horizon when demand response events are activated. Results show that deep learning models have a higher ability to provide accurate estimates of baselines than other models. However, forecasting the energy reductions of dwelling units when a demand response event is activated within a very short horizon, remains a challenge for all models.