To facilitate the co-design of next generation hardware architectures, it is critical to characterize the workloads of deep learning (DL) applications and assess their computational patterns on different levels of the execution stack. Time series prediction is one such DL application heavily used in areas that include critical decision making: ensuring power grid resiliency, climate forecasting, transportation infrastructure optimization, stock market prediction, etc. Unlike cross-sectional data (e.g. images), time-series data is inherently sequential, posing challenges to parallelization in the context of deep learning.

Residential buildings account for a significant proportion of overall energy consumption across the world. Decentralized room level Air Conditioners (ACs) are a commonplace in developing countries such as India, contributing a major share (34% in India) of the total residential energy consumption. Option to independently control each AC presents a prime opportunity for an energy saving system. Thus, we propose PACMAN to non-intrusively (using only the temperature information) predict AC energy consumption prior to usage and estimate energy consumption post-usage. We discuss various possible applications and use cases of such feedback for the occupants. To empirically validate the performance of PACMAN, we conducted an in-situ study across seven homes in Delhi (India). We collected around 2200 hours of usage data from the different ACs, room types, and thermostat temperatures. We achieved an average accuracy of 85.3% and 83.7% with the best accuracy of 97.0% and 93.3% for the estimation and prediction of AC energy consumption respectively, across all homes. Towards the end, we discuss various outlier scenarios, opening up multiple directions for further research in this domain.