Energy Consumer Model Estimation Using Machine Learning
A novel utility model of the daily electricity consumption patterns of any residential unit is proposed. The model is semi-parametric – its internal parameters are tractable to straightforward interpretations from the standpoints of microeconomics as well as game theory. The model considers two categories of appliances, fixed and shiftable loads. A modified Cobb-Douglas preference function is applied to incorporate utility tradeoffs in operating shiftable loads at different hours of the day, based on the consumer’s intrinsic preference and time-of-use pricing. The task of model parameter estimation is formulated in terms of constrained optimization, with an L2 error-based, regularized loss function, and ϵ-insensitive error constraints. It has been theoretically proven that there is a unique global optimum. A dual gradient ascent learning algorithm is developed to obtain model parameters that best fit real time series data of the unit’s daily shiftable and fixed loads. For a unit without sensors to separately monitor each load category, an unsupervised learning algorithm for partial load disaggregation is developed, which hybridizes the Gaussian mixture model with non-negative matrix factorization. Simulations with real data of two units highlight the effectiveness of the proposed methodology.