In today's dynamic energy landscape, Demand Response (DR) programs have become essential for ensuring grid stability and operational efficiency. These programs empower prosumers to adjust their electricity consumption collectively or voluntarily, often against a predefined baseline load. Accurate estimation of this flexibility is pivotal for the success of DR programs, particularly in scenarios where only aggregated building data is available. This study highlights the significance of constructing a building's flexibility baseline load, which approximates its electricity consumption and production in the absence of specific DR events. The deviation of energy consumption and production from this baseline, by predetermined margins, signifies potential flexibility. However, accurately forecasting next-day load variations is essential for evaluating flexibility spread around the baseline. In addressing this need, we employ Long Short-Term Memory (LSTM) networks, optimizing their parameters to significantly enhance performance in predicting energy consumption and production for all weekdays. To validate our proposed solution, we conduct real-world experiments using prosumer data collected from a building in Austria.