Approximating discrete probability distributions with dependence trees
IEEE Transactions on Information Theory
| Hi-index | 0.00 |
The ongoing smart grid transformation in utility networks is making available fine grained measurements of electricity consumption. To realize the full potential of the collected data we apply sophisticated data analytics and machine learning techniques to correlate consumption with other types of demographic data (household surveys and tax records) to place the collected consumption data within the right context. This context setting is achieved by a rigorous feature selection procedure, followed by clustering to group customers into peer groups. The statistical information gleaned from these peer groups are then used to identify outliers and define new services both for the utility (energy audits) and the end consumer (Home Energy Health Management systems). Analysis shows that outlier detection within clusters is better able to target customers than outlier detection without clustering: on average, half the outliers found in the clusters would not be outliers in the overall population. The techniques employed could also be used to detect anomalous usage patterns that may be indicative of fraudulent use of electricity.