In the event of a severe accident in nuclear power plants (NPPs), an important issue is the hydrogen generation due to the oxidation of the fuel cladding at high temperatures inside the reactor as the coolant disappears and the core melts. During normal operation, the hydrogen concentration in containment should be kept below 4%. However, if the hydrogen concentration increases above 10% or more during a severe accident, explosive combustion reaction leading to detonation may occur and eventually it can lead to damage to the containment. Therefore, it is important to predict the hydrogen concentration in severe accidents. There have been several studies by researchers to predict the hydrogen concentration in containment by using many artificial-intelligence (AI) techniques such as fuzzy neural network (FNN) and cascaded fuzzy neural network (CFNN). This study suggests the prediction of hydrogen concentration in containment under severe accidents using a deep neural network (DNN) method. Since the severe accident data cannot be obtained from actual NPPs, we verified the proposed method based on simulation data acquired using the modular accident analysis program (MAAP) code. The DNN model shows excellent prediction performance when a variety of loss of coolant accident (LOCA) data is applied. The proposed DNN model allows operators to predict the exact hydrogen concentration in containment at the beginning of the accident. Prediction of this hydrogen concentration will help to ensure safety by reducing the risk of the hydrogen combustion and explosion in a containment.