Optimising multi-factor assistance in a deep learning-based electricity forecasting model with climate resilience: an Australian case study

With the increasing incidence of power blackouts attributed to climate change, climate-resilient load forecasting is increasingly necessary to enable timely network reconfiguration during extreme weather events. This paper proposes a generalised multi-factor Deep Learning (DL) model to forecast electricity load in Distribution Networks (DNs) during extreme climate periods. We optimise factors that affect forecast accuracy, including input matrix structures, calendar effects, and correlation-based preceding temperatures. The novel input feature selection decreases the Mean Absolute Percentage Error (MAPE) by 30.73% when compared to using instantaneous temperature alone. The model has been developed based on three real DNs in Victoria, Australia, during the wildfire seasons of 2015-2020. The sensitivity to large- scale climate variability, such as El Niño or La Niña, has also been assessed, showing minimal impact. Our model achieves an average forecast MAPE of 2.95%, ensuring that reconfiguration events mitigate rather than exasperate high load conditions during periods of extreme stress.