Day ahead load forecasting for the modern distribution network-A Tasmanian case study

Abstract

Penetration of distributed energy resources in distribution networks is predicted to increase dramatically in the next seven years, bringing with it the opportunity for utilities to have a greater presence at low levels of the network. To achieve this effectively, utilities will require accurate short term load forecasts. This paper presents a novel neural network-based load forecasting system that applies recent advances in neural attention mechanisms. The forecasting system is trained and assessed on ten years of historical half-hourly load, weather, and calendar data to produce a 24-hour horizon half-hourly online forecast. When forecasting during anomalous peak holiday periods on a feeder that has a typical load of less than 1000kVA the forecasting system achieves a MAPE of 7.4% and a mean error of -15kVA. The forecasting system is implemented in a residential battery trial and is able to successfully forecast major peaks with sufficient lead time and accuracy to enable the fleet of batteries to charge ahead of time and provide network support.