The electrification of heating and transportation, together with the increasing penetration of distributed renewable energy resources, is significantly transforming load dynamics in low-voltage distribution grids. Although forecasting models are typically evaluated using statistical error metrics, the physical implications of forecast inaccuracies on grid operation remain insufficiently understood. This study investigates how node-level net load forecasting errors propagate to grid-state variables, namely line loading and voltage deviations. Using a real German low-voltage grid topology and customer-level data, cluster-based global forecasting architectures employing machine learning and deep learning are coupled with non-linear power flow simulations to quantify the operational impact of forecast inaccuracies. The results show that improvements in forecasting accuracy translate almost proportionally into reductions in line-loading and voltage-deviation errors. In particular, models that accurately capture net load valleys during periods of high PV generation also achieve superior grid-state prediction performance, highlighting the operational relevance of valley-oriented forecasting metrics. By linking predictive performance directly to physical grid constraints, the proposed framework provides a systematic method for assessing the operational relevance of forecasting models. Beyond exploratory analysis, the study further derives explicit relationships that enable system operators to translate forecasting error metrics into expected grid-state deviations under comparable network conditions.
Papadopoulos et al. (Mon,) studied this question.