【Advances in Applied Energy】明天要耗多少电？一个精准、自适应、且可解释的理论指导短期电力负荷预测模型

（1） Constructed a novel knowledge and data dual-driven approach (Adaptive-TgDLF) that makes full use of human knowledge and advanced deep learning techniques .

（2） Employed adaptive learning to utilize load data at various locations and times, which improves the generalization ability of model .

（3） Proposed a method to mine interpretability of the deep-learning model for load forecasting via attention matrix .

（4） The proposed model is stronger (being 16% more accurate), more robust (the performance of the proposed model with 50% weather noise is the same as that of the previous efficient model without weather noise), easier to train (saving more than half of the training time), and requires less data .

Electrical energy is essential in today’s society. Accurate electrical load forecasting is beneficial for better scheduling of electricity generation and saving electrical energy. In this paper, we propose an adaptive deep-learning load forecasting framework by integrating Transformer and domain knowledge (Adaptive-TgDLF). Adaptive-TgDLF introduces the deep-learning model Transformer and adaptive learning methods (including transfer learning for different locations and online learning for different time periods), which captures the long-term dependency of the load series, and is more appropriate for realistic scenarios with scarce samples and variable data distributions.Under the theory-guided framework, the electrical load is divided into dimensionless trends and local fluctuations. The dimensionless trends are considered as the inherent pattern of the load, and the local fluctuations are considered to be determined by the external driving forces. Adaptive learning can cope with the change of load in location and time, and can make full use of load data at different locations and times to train a more efficient model. Cross-validation experiments on different districts show that Adaptive-TgDLF is approximately 16% more accurate than the previous TgDLF model and saves more than half of the training time. Adaptive-TgDLF with 50% weather noise has the same accuracy as the previous TgDLF model without noise, which proves its robustness.We also preliminarily mine the interpretability of Transformer in Adaptive-TgDLF, which may provide future potential for better theory guidance. Furthermore, experiments demonstrate that transfer learning can accelerate convergence of the model in half the number of training epochs and achieve better performance, and online learning enables the model to achieve better results on the changing load .

Keywords

load forecasting; deep-learning; domain knowledge; transfer learning; online learning; interpretability

图1：Adaptive-TgDLF示意图，包含由领域知识确定的无量纲趋势以及由Transformer模型预测的局部波动，以及使用自适应学习去应对负荷在空间和时间上的变化

图2：丰台区负荷真实值（黑色）与Adaptive-TgDLF预测值（红线）对比图

图3：（房山，门头沟）区域组迁移学习效果图，蓝色实线代表非迁移学习，其他颜色实线代表采用不同策略的迁移学习

图4：丰台、平谷和房山区域在线学习效果图