NEURAL-NETWORK-BASED SUBSTATION SHORT-TERM LOAD FORECASTING

There are many algorithms reported in the literature to forecast the total real load of a power system. But in a power system, the local area loads (both real and reactive loads) are more helpful for dispatching center operators to schedule generation outputs. An approach to substation load (both real and reactive power) forecast by an artificial neural network (ANN) is presented in this paper Characteristic data of substation load collected continuously by the Supervisory Control and Data Acquisition (SCADA) system of the dispatch center are used for the forecast. The characteristic data include substation historical loads, ambient temperature, relative humidity, system frequency, substation voltages, shunt capacitor status and transformer tap ratios. Since the forecast is based on data acquired by SCADA, the time interval between data samples can be as short as minutes or even seconds; thus, the forecasted load model is suitable for dynamic load studies. Furthermore, the algorithm to vary the number of hidden units is applied to this research and makes it no longer necessary to pre-determine the number of hidden units. To speed up the training process, an adaptive training process of ANN is also Taiwan Power Company system to forecast both real and reactive loads and the testing results are satisfactory.