The transformation of the energy system towards volatile renewable generation, increases the need to manage decentralized flexibilities more efficiently. For this, precise forecasting of uncontrollable electrical load is key. Although there is an abundance of studies presenting innovative individual methods for load forecasting, comprehensive comparisons of popular methods are hard to come across.In this paper, eight methods for day-ahead forecasts of supermarket, school and residential electrical load on the level of individual buildings are compared. The compared algorithms came from machine learning and statistics and a median ensemble combining the individual forecasts was used.In our examination, nearly all the studied methods improved forecasting accuracy compared to the naïve seasonal benchmark approach. The forecast error could be reduced by up to 35% compared to the benchmark. From the individual methods, the neural networks achieved the best results for the school and supermarket buildings, whereas the k-nearest-neighbor regression had the lowest forecasting error for households. The median ensemble narrowly yielded a lower forecast error than all individual methods for the residential and school category and was only outperformed by a neural network for the supermarket data. However, this slight increase in performance came at the cost of a significantly increased computation time. Overall, identifying a single best method remains a challenge specific to the forecasting task.

中文翻译：

---

不同建筑类型短期用电负荷预测方法比较

能源系统向不稳定的可再生能源转型，增加了更有效地管理分散灵活性的需求。为此，对不可控电力负荷的精确预测是关键。虽然有大量的研究提出了创新的个体负荷预测方法，但很难对流行的方法进行综合比较。建筑物进行比较。比较算法来自机器学习和统计学，并使用了结合各个预测的中值集合。在我们的检查中，与朴素的季节性基准方法相比，几乎所有研究的方法都提高了预测准确性。与基准相比，预测误差最多可减少 35%。从个别方法来看，神经网络对学校和超市建筑取得了最好的结果，而 k-最近邻回归对家庭的预测误差最低。与住宅和学校类别的所有单独方法相比，中值集合的预测误差略低，并且仅优于超市数据的神经网络。然而，这种性能的轻微提升是以显着增加计算时间为代价的。总的来说，确定单一的最佳方法仍然是预测任务特有的挑战。而 k-最近邻回归对家庭的预测误差最低。与住宅和学校类别的所有单独方法相比，中值集合的预测误差略低，并且仅优于超市数据的神经网络。然而，这种性能的轻微提升是以显着增加计算时间为代价的。总的来说，确定单一的最佳方法仍然是预测任务特有的挑战。而 k-最近邻回归对家庭的预测误差最低。与住宅和学校类别的所有单独方法相比，中值集合的预测误差略低，并且仅优于超市数据的神经网络。然而，这种性能的轻微提升是以显着增加计算时间为代价的。总的来说，确定单一的最佳方法仍然是预测任务特有的挑战。性能的这种轻微提升是以显着增加计算时间为代价的。总的来说，确定单一的最佳方法仍然是预测任务特有的挑战。性能的这种轻微提升是以显着增加计算时间为代价的。总的来说，确定单一的最佳方法仍然是预测任务特有的挑战。