Improving forecast accuracy for grid demand and renewables supply with pattern-match features

Short term forecasts of electricity demand and renewables supply are vital for smart grid applications, and in particular for the successful integration of renewables. Improved forecasting of these quantities supports several aspects of smart grid control and deployment, including planning, pricing, staffing, and more. Globally, there is a target for renewables engagement of ∼36% (percentage of renewables in the global energy mix) by 2030. This is expected to increase global GDP by 1.1%, and to substantially reduce GHG emissions. It is well known that one of the main barriers to the integration of renewables is the relatively unpredictable nature of renewables availability. A further difficulty (in many grid applications, irrespective of renewables), is the similarly unpredictable nature of energy demand. In this paper we explore an idea in the area of time series prediction that was developed specifically for grid-relevant applications, and we test this method on wind-speed data and on energy demand data. The basis of the approach is the exploitation of patterns in the time series that are maximally similar to the current context. However, unlike previous work that has explored the use of such pattern-matching directly for prediction, we instead use the outcome of pattern matching to generate one or more additional features for the subsequent machine learning. Experiments suggest that (i) the approach reliably leads to improved forecasts; and (ii) the approach has wide applicability, independent of the machine learning approach in use.