The purpose of this paper is to compare the performance of different structures of Artificial Neural Networks (ANNs) regarding the output variables used for short term forecasting of the next day load of the interconnected Greek power system. In all cases the output variables are the hourly actual loads of the next day. The classical ANN design adopts an ANN model with 24 output variables. Alternatively, 24 different ANN models can be implemented for each hour of the day. This solution can affect the selection of input variables indirectly. In this paper, various scenarios of the solution of 24 different ANN models are going to be studied with different sets of input variables using the scaled conjugate gradient training algorithm, for which a calibration process is conducted regarding the crucial parameters values, such as the number of neurons, the type of activation functions, etc. The performance of each structure is evaluated by the Mean Absolute Percentage Error (MAPE) between the experimental measurements and estimated values of the hourly load demand of the next day for the evaluation set in order to specify the optimal ANN. Next, the load demand for the next day of the test set (with the historical data of the current year) is estimated using the best ANN structure, to verify the behaviour of ANN load prediction techniques. Finally the classical design and different proposed structures are compared.