The version presented here may differ from the published version or from the version of the record. Please see the repository URL above for details on accessing the published version and note that access may require a subscription.

Technological advancements in the field of electrical energy distribution and utilization are revolutionizing the way consumers and utility providers interact. In addition to allowing utility companies to monitor the status of their network independently in autonomous fashion, data collected by smart meters as part of the wider advanced metering infrastructure, can be valuable for third parties, such as government authorities. The availability of the information, the granularity of the data, and the real-time nature of the smart meter, means that predictive analytics can be employed to profile consumers with high accuracy and approximate, for example, the number of individuals living in a house, the type of appliances being used, or the duration of occupancy, to name but a few applications. This paper presents a machine learning model comparison for unemployment prediction of single household occupants, based on features extracted from smart meter electricity readings. A number of nonlinear classifiers are compared, and benchmarked against a generalized linear model, and the results presented. To ensure the robustness of the classifiers, we use repeated cross validation. The results revealed that it is possible to predict employability status with Area Under Curve (AUC) = 74%, Sensitivity (SE) = 54% and Specificity (SP) = 83%, using a multilayer perceptron neural network with dropout, closely followed by the results produced by a distance weighted discrimination with polynomial kernel model. This shows the potential of using the smart metering infrastructure to provide additional autonomous services, such as unemployment detection, for governments using data collected from an advanced and distributed Internet of Things (IoT) sensor network.

No abstract