Abstract :
Load prediction is an important tool for grid utilities and power companies for managing the power system. This has traditionally mainly been applied at transmission and subtransmission system levels. However, as the traditional grid evolves into a smart grid, load prediction at smaller scales becomes necessary for efficient management and operation. In this paper we investigate how the qualitative and statistical properties of load time series change as a function of the number of individual loads aggregated in the series, and how these properties influence the predictability of the time series. We study the performance of a traditional autoregressive model, a wavelet-based model, an Echo State Network, and a variation of Case-Based Reasoning (CBR) at the subtransmission (~10000 customers), distribution substation (~150 customers) and single-meter level. For all the four prediction methods, we employ an evolutionary algorithm as a meta-learner to automatically optimize the free parameters for each model-dataset combination. We find that relatively accurate predictions can be made at finer granularity, but care must be taken in choosing, tuning and analyzing the prediction model as the regularity of the consumption patterns decreases.