Real-Time Stream Mining Electric Power Consumption Data Using Hoeffding Tree with Shadow Features

摘要

Many energy load forecasting models have been established from batch-based supervised learning models where the whole data must be loaded to learn. Due to the sheer volumes of the accumulated consumption data which arrive in the form of continuous data streams, such batch-mode learning requires a very long time to rebuild the model. Incremental learning, on the other hand, is an alternative for online learning and prediction which learns the data stream in segments. However, it is known that its prediction performance falls short when compared to batch learning. In this paper, we propose a novel approach called Shadow Features (SF) which offer extra dimensions of information about the data streams. SF are relatively easy to compute, suitable for lightweight online stream mining.