Data science from Oracle Utilities unlocks customer usage information from Advance Metering Infrastructure (AMI) data. These insights are discovered by sophisticated machine learning models and are distinct from statistical regressions or rules-based analysis. Three separate and important questions can be answered with our models: Presence Discovery, Detection and Disaggregation.

• Presence Discovery – Predicting existence of a large appliance or EV at a household is referred to as presence discovery. The output of our models is typically expressed as a confidence level. The results of our models can be used to infer presence on unknown households, which are households the model has not seen during training. These results can be augmented by surveys such as Opower Home Energy Audit (HEA).
• Detection - Predicting when a large appliance is in use or EV is charging at 15min/30min/1hour (AMI) granularity is referred as detection. This information is of particular importance as input to energy efficiency saving programs with actionable insights within a time of use or demand charge tariff structure.
• Disaggregation - Predicting energy consumption of large appliances and EVs at matching AMI granularities is referred as disaggregation.

Deep learning models

Based on years of research and experience, Oracle's data scientists determined that a deep learning methodology is best suited for presence discovery, detection and disaggregation. Deep learning is at the forefront of machine learning and is particularly well suited for analysis of patterns within time series, which is specifically the problem presented by AMI data. Oracle currently has two patents1 on presence discovery and disaggregation and three patents pending in the field of deep learning--a testament to the level of the sophistication our models.

Our suite of models is based on deep learning innovations and provide very accurate insights based on AMI data (including the date and time) and weather. Our models operate on 15 min, 30 min and 60 min AMI interval data by discovering, detecting and disaggregating all make and models of electric vehicles (EVs), HVAC, central air conditioning unit, furnace, heat pumps, refrigerator, clothes dryer, clothes washer, oven, dishwasher, electric water heaters, pool pumps and more. Oracle’s deep learning models compare favourably to the published results in the scientific literature through the implementation of more than a dozen objective metrics. Our internal tests only use unseen homes which are homes that are not part of the training but have submetered data. The results presented in this document are from our first-generation models. Our expectation is that new, labelled training data will continue to improve our models.