A deep-learning based few-shot object detection algorithm is adopted to perform cable damage detection. Cross domain knowledge transfer is exploit so that the deep learning model can learn to identify cable damage by given a few samples.

In this paper, we consider the load forecasting for a new user in the system by observing only few shots (data points) of its energy consumption. We propose to utilize clustering to mitigate the challenges brought by the limited samples. Specifically, we first design a feature extraction clustering method for categorizing the historical data. Then, the load forecast for new users is conducted through a two-phase Long Short Term Memory (LSTM) model, which inherits prior knowledge from the clustering results.

申请公开了基于聚类的小样本负荷预测方法和预测装置，其中预测方法包括：对历史电力负荷和待预测电力负荷进行特征提取，获得特征向量；根据得到特征向量将历史电力负荷和待预测电力负荷集成聚类，得到聚类结果；采用小波降噪算法对聚类结果进行降噪，并将降噪后的数据进行均值化处理，得到预设长度的时序数据；其中预设长度的时序数据包括历史时序数据和待预测时序数据；将预设长度的时序数据输入二阶长短期记忆神经网络，得到电力负荷的预测结果；其中历史时序数据用于训练二阶长短期记忆神经网络，待预测时序数据用于调整训练完成后的二阶长短期记忆神经网络。本申请在待预测电力负荷稀缺情况下仍然有优秀预测表现。