Iterative memory-driven load forecast network model for accuracy improvement

Abstract

High penetration of renewable energy generation and intensive application of new-type load control measures in power system greatly increase the difficulty of load forecast (LF). Long-term dependencies in load series limit LF accuracy improvement of back propagation neural network (BPNN) and its combination variants. Inspired by coexisting features of long-term trend and non-trend fluctuation in load series, a novel iterative memory-driven load forecast network (IMDLFN) model is proposed to forecast system-level load in the scenario. It creatively adopts memory-driven units to convey long-term trend information learned from the previous time steps and to autonomously balance long-term memory and short-term memory, which effectively deals with long-term dependencies and perfectly circumvents weight selection problem of BPNN combination variants. LF for municipal power system in China demonstrates that the proposed model can remarkably improve LF accuracy especially for trend/non-trend coexisting load series.

Keywords

• Load forecast
• Load control
• Renewable energy generation
• Power systems