Inadequacy of Standard Algorithms and Metrics for Short-Term Load Forecasts in Low-Voltage Grids

Short-term load forecasting is becoming popular in low-voltage grids due to the availability of Smart Meter (SM) measurements. At the individual customer level, load profiles are characterized by a high volatility and are hard to predict, i.e., traditional forecasting algorithms generally perform poorly. Furthermore, their evaluation based on point-wise metrics like the MAPE and RMSE give a biased image of the actual prediction accuracy. In this paper, we discuss the performance of four forecasting methods, including a novel algorithm based on Markov chains. We investigate different approaches to assess the forecast performance on volatile loads and we notably compare the statistical properties of the predictions. The analysis highlights the inadequacy of traditional point forecasting algorithms and standard point-wise metrics.