Short-Term Electricity Load Forecasting for Fine-Grained Data with PLAM

In this section, we address Short-Term Load Forecasting (STLF) of half-hourly electricity load for one day ahead as it is vital for the efficient operation of electric power systems and grid management. While electricity load forecasting at the aggregate level across many households has been extensively studied, electrical load forecasting at fine-grained geographical scales of households, which is the case studied in this chapter, is more difficult as we move toward lower levels of load aggregation, since local demand profiles are highly volatile and noisy. There is a rich literature on forecasting the average aggregated electricity load, and we have already seen in previous sections of this chapter that the use of semi-parametric generalized additive models (GAM), when forecasting is addressed at higher levels of aggregation, is accurate and flexible, since the aggregated load pattern contains relatively smooth additive components. However, lower aggregation levels result in high-resolution data that is highly volatile, and forecasting the average load using GAM models with smooth components does not provide meaningful information about the future demand. To enhance the forecast accuracy, we need to incorporate some much more irregular and volatile effects. We focus on the analysis of such hybrid additive models applied on smart meters data and compare it to the forecasting performances of classical additive models at low aggregation levels. Exploiting a unified basis representation of the nonlinear additive components allows us to use some automatic variable selection and estimation procedures for sparse high-dimensional Partially Linear Additive Models (PLAMs) with nonparametric additive components of heterogeneous smoothness, which will be efficiently applied to forecast electricity data whatever their level of aggregation is. The results and discussions in this section follow closely a recent work (Amato et al., International Journal of Forecasting, 37:171–185, 2020) performed by some of the authors.