A cloud-based building automation system was developed, and implemented in an actual building featuring a thermo active building system, ground source heat pump system, and battery. Based on the operation and measurement data stored by this system, a predictive model was created using a neural network. In the hyper-parameters of the neural network, the number of hidden layers, nodes in each hidden layer, and history levels of the input data were optimized by particle swarm optimization with the mutation method. Models that predict future behavior of heating, ventilation, and air-conditioning system were evaluated according to normalized mean absolute error, with the best model obtaining 0.0796.