A New Forecasting Algorithm Based on Neighbors for Streaming Electricity Time Series

摘要

This work presents a new forecasting algorithm for streaming electricity time series. This algorithm is based on a combination of the K-means clustering algorithm along with both the Naive Bayes classifier and the K nearest neighbors algorithm for regression. In its offline phase it firstly divide data into clusters. Then, the nearest neighbors algorithm is applied for each cluster producing a list of trained regression models, one per each cluster. Finally, a Naive Bayes classifier is trained for predicting the cluster label of an instance using as training the cluster assignments previously generated by K-means. The algorithm is able to be updated incrementally for online learning from data streams. The proposed algorithm has been tested using electricity consumption with a granularity of 10 min for 4-h-ahead predicting. Our algorithm widely overcame other four well-known effective online learners used as benchmark algorithms, achieving the smallest error.