Energy Anomaly Detection and Modelling on Smart Premises using SDAR

Energy as a commodity is facing a globally prevalent shortage with the ever increasing gap between the demand and supply. The strain on the non-renewable conventional energy resources is evident from the multitude of industries in the manufacturing hubs of China that have come to standstill in the face of non-availability of coal. The phenomenon is particularly acute in all developing countries with an unbalanced approach to the energy management vis-a-vis distribution losses. The Internet of Things (IoT) ecosystem enabled smart homes and industrial premises have shown promising application towards achieving energy efficiency via creation of dynamically adjusting demand based systems. It is possible to predict the energy consumption requirements based on application of Machine Learning (ML) based models on statistical data obtained from energy sensors. The data streams from IoT devices however, more often than not throw up surprises in the form of outliers and changes which can affect the Machine learning based time series forecasting. The problem is more accentuated in the case of non-stationary time series sources where it is imperative to ascertain whether an anomaly is momentarily affecting the time series as an outlier or it is a permanent change never returning to the original trend. This paper uses a sequentially discounting auto regression (SDAR) learning algorithm to detect and classify the anomalies in energy consumption usage for model accuracy. Specifically, we have applied the online SDAR algorithm on energy consumption IoT dataset from kaggle as a demonstration to distinguish between outliers and permanent changes over the time series which can be used for interpretability and increasing model accuracy while prediction of energy consumption. We were able to forecast sudden changes in the energy consumption requirements well in advance, based on the previous years' usage patterns as the results indicate.