How Much Training Data Is Enough? An Empirical Study on Load Forecasting for a Distribution Transformer in Germany

The increasing digitalization of distribution grids and the gradual deployment of smart monitoring infrastructure are beginning to generate valuable data streams at the lowvoltage level. These advances enable data-driven forecasting and the integration of flexibility to enhance grid operation. However, since monitoring equipment has only recently been installed, historical data availability remains limited. This raises a central question: how much training history is necessary to achieve reliable day-ahead load forecasts that can support flexibility use in distribution grids? This paper investigates this issue by systematically examining the influence of training dataset length on forecasting performance for a low-voltage transformer in Germany. A broad set of machine learning models, including linear approaches, tree-based ensembles, and deep learning methods, was evaluated under a day-ahead forecasting framework. The results show that tree-based ensembles consistently outperform other models, with CatBoost delivering the most accurate and stable forecasts across all training horizons. Yet, when the training data spans only up to three months, simple naive baselines remain highly competitive due to their zero training cost. From the fourth month onward, machine learning models increasingly leverage longer training histories to decisively surpass naive methods. These findings provide practical guidance for distribution system operators: employ naive averages in the early stages of monitoring and transition to advanced ensemble methods once at least four months of data are available.