Autoregressive model for individual consumption data - Sparsity recovery and significance test

Understanding consumer flexibility and behavior patterns is becoming increasingly vital to the design of robust and efficient energy saving programs. Accurate prediction of consumption is a key part to this understanding. Existing prediction methods usually have high relative errors that can be larger than 30%. In this paper, we explore sparsity in users' past data and relationship between different users to increase prediction accuracy. We show that using LASSO and Granger Causality techniques, prediction accuracy can be significantly improved in comparison to existing algorithms. We use mean absolute percentage error (MAPE) as the criteria.