The application of artificial neural networks or other techniques in load forecasting usually outputs quality results in normal conditions. However, in real-world practice, a remarkable number of abnormalities may arise. Among them, the most common are the historical data bugs (due to SCADA or recording failure), anomalous behavior (like holidays or atypical days), sudden scale or shape changes following switching operations, and consumption habits modifications in the face of energy price amendments. Each of these items is a potential factor of forecasting performance degradation. This work describes the procedures implemented to avoid the performance degradation under such conditions. The proposed techniques are illustrated with real data examples of current, active, and reactive power forecasting at the primary substation level.