Quality prediction of load evolution at different levels of distribution network is a basic requirement for adequate operation planning of modern power systems. This paper describes the models, based on artificial neural networks, developed for active and reactive power forecasting at primary substations' transformers. The main goal consists on defining a regression process characterized by good quality estimates of those future values, based on historical data. Anticipation interval shall include from the next hour to one week in advance. The implemented forecasting tool is able to deal with noisy data, holidays and special occasions and adapts forecasts in case of power network reconfiguration whenever planned. Used techniques and implementation foundations of selected forecasting models are reported. Finally, the potential of the adopted approach is sustained by illustrative examples.