Analysis of Electricity Usage Time Series with K-means Clustering

Sammanfattning: As the amount of collected and analysed data for electricity usage from buildings is increasing it becomes an important component in energy efficiency efforts. A huge problem when developing algorithms fordetection of anomalous electricity usage series is the lack of data setswith annotated normal usage series for training and evaluation purposes.The aim was thus to investigate if there is a typical common usagepattern in electricity usage series coming from offices and multi-family residential buildings with the unsupervised machine learning method K- means clustering. If common usage patterns are found they couldpotentially be used as a reference to what normal usage is in anomaly detection algorithms. Another aim was also to investigate how theclustering is affected by different time resolutions and normalization methods. The tested normalization methods were Min-max normalization, Z- score normalization and Robust normalization. To select the optimalnumber of clusters for the K-means clustering method and to validate the clustering Davies-Bouldin index, Silhouette index and Calinski-Harabasz index were used. The results showed that clustering with Robust normalization produced the best validity indices. However the resulting K-means clusteringoften allocated the majority of the series to one cluster and a few outliers in another. Min-max normalization performed the second best according to the validity indices and this normalization method produced clusters where most offices were allocated to one cluster and most residential series to another. Using daily resolution data instead of hourly did not change the result significantly. From the results it could be concluded that offices and residential buildings have to a large extent distinctly different electricity usage patterns which potentially can be utilized in anomaly detection algorithms. Also the results showed the importance of choosing normalization method with care as it could affect the resulting clustering hugely.