Reporting Interval Impact on Deep Residential Energy Measurement Prediction

Forecasting and anomaly detection for energy time series is emerging as an important application area for computational intelligence and learning algorithms. The training of robust data-driven models relies on large measurement datasets sampled at ever increasing rates. Thus, they demand large computational and storage resources for off-line power quality analysis and for on-line control in energy management schemes. We analyze the impact of the reporting interval of energy measurements on deep learning based forecasting models in a residential scenario. The work is also motivated by the development of embedded energy gateways for online inference and anomaly detection that avoid the dependence on costly, high-latency, cloud systems for data storage and algorithm evaluation. This, in turn, requires increased local computation and memory requirements to generate predictions within the control sampling period. We report quantitative forecasting metrics to establish an empirical trade-off between reporting interval and model accuracy. Additional results consider the time scale variable feature extraction using a time series data mining algorithm for multi-scale analytics.