Growing electrical demand on the electric system along with the rising use of renewable energy sources is highlighting the importance of energy flexibility management on the electric grid. The Electric System Operators at both transmission (TSO) and distribution level (DSO) are responsible to ensure the security of supply and efficiency of the grid under strict balancing conditions (demand equals supply at every time instant). Acting on both generation and demand to maintain this equilibrium considering the technical constraints of the grid is known as flexibility management and it requires accurate generation and demand forecasting to predict possible critical events and react accordingly. The objective of this paper is to analyze the performance of different forecasting methods for predicting demand at the substation level. Substation level data is the result of aggregating the consumption and generation data of multiple points on the grid. Results show that current state of the art algorithms, such as deep learning models, perform better than simpler methods, such as random forests, specially when datasets do not present clearly repetitive profiles. Deep learning models manage to reduce forecasting error by 16% on average compared to random forest models on next day load forecasting, whereas the forecasting error reduction on next hour load forecasting is 5%.