Efficient clustering for aggregate loads: An unsupervised pretraining based method

Load management is an important issue for electricity system stability and renewable energy application. Load clustering is a key topic of load management. However, at most of the time load distribution is complex and is highly related to wide socioeconomic and demographic factors. This makes load clustering a hard problem especially when only aggregate load data is available. This paper proposes a method that firstly encodes an arbitrary load into an embedding centroid vector, and secondly carries out clustering based on the embeddings. An unsupervised pretraining approach is proposed as an embedding system. In this framework, only aggregate load data is needed. This paper put forward a metric to identify the accuracy of the embedding system. Under this metric, the proposed method is superior to a naïve auto-encoding approach which is a successful unsupervised pretraining method in data compression and reconstruction. When an outer dataset is applied, the proposed method can still get higher scores, which indicates good generalization ability of the method. Results show that the embedding centroids have a better clustering tendency than conventional features. In the clustering based on the embedding centroids, not only daily patterns but also monthly patterns are captured by the method.