A novel application of machine learning techniques for activity-based load disaggregation in rural off-grid, isolated solar systems

In power systems and electricity markets, accurate monitoring and prediction of electricity demand is important in order to manage real-time load balancing and for distribution and transmission planning. In the context of rural electrification, the uncertainty of both supply — often with 100% renewables — and demand is large and the central reason that solar-based micro-grids and home systems can be expensive. Oversizing battery storage and solar panel size or limiting the electricity available to each user can alleviate this uncertainty but increases system costs. Historical data for solar irradiance is available and allows quantification of the uncertainty associated with supply. However, electricity demand in rural households is not well characterized in developing world contexts. In an ideal scenario, the data should be captured on a per-household level and must be measured on a sufficiently fine time resolution to translate the demand onto individual, user activities. This massive quantity of household demand data requires automatic tools to convert time series power usage data into data on the use of individual appliances. This paper presents a methodology, based on data acquired in individual, isolated solar home systems in Jharkhand, India, on utilizing classification and clustering algorithms to create activity-based models that can be used to conduct load forecasts. Additional statistical data analysis can yield insights on users' power consumption behavior in relation to exogenous variables such as time of day and conditioned on ambient air temperature.