Short-term load forecasting (STLF) is vital for the daily operation of power grids. However, the non-linearity, non-stationarity, and randomness characterizing electricity demand time series renders STLF a challenging task. To that end, different forecasting methods have been proposed in the literature for day-ahead load forecasting, including a variety of deep learning models that are currently considered to achieve state-of-the-art performance. In order to compare the accuracy of such models, we focus on national net aggregated STLF and examine well-established autoregressive neural networks of indicative architectures, namely multi-layer perceptrons, N-BEATS, long short-term memory neural networks, and temporal convolutional networks, for the case of Portugal. To investigate the factors that affect the performance of each model and identify the most appropriate per case, we also conduct a post-hoc analysis, correlating forecast errors with key calendar and weather features. Our results indicate that N-BEATS consistently outperforms the rest of the examined deep learning models. Additionally, we find that external factors can significantly impact accuracy, affecting both the actual and relative performance of the models.

翻译：短期负荷预测（STLF）对电网日常运营至关重要。然而，电力需求时间序列的非线性、非平稳性和随机性特征，使STLF成为一项具有挑战性的任务。为此，文献中提出了多种日前负荷预测方法，其中包括当前被认为实现最优性能的各类深度学习模型。为比较这些模型的预测精度，我们聚焦国家净聚合的短期负荷预测，针对葡萄牙案例，检验了具有代表性架构的成熟自回归神经网络，具体包括多层感知机、N-BEATS、长短期记忆神经网络及时间卷积网络。为探究影响各模型表现的因素并确定每类场景最适用的模型，我们还开展了事后分析，将预测误差与关键日历特征和天气特征进行关联。结果表明，N-BEATS在性能上持续优于其他被检验的深度学习模型。此外，我们发现外部因素可显著影响预测精度，既改变模型的绝对性能，也改变其相对性能。