Learning for Interval Prediction of Electricity Demand: A Cluster-based Bootstrapping Approach