Volume 54

Autoregressive Transformer for Predicting and Synthesizing Residential Load Kosaku Nakano, Daishi Sagawa, Shingo Ito, Keisuke Ando, Ryohta Kitagawa, Kenji Tanaka

Abstract

Accurate load forecasting is crucial for efficient energy management, particularly due to the increasing adoption of renewable energy sources and the growing need for grid stability. This study introduces an autoregressive transformer encoder-decoder model that advances residential electricity load prediction and synthesis. Unlike the conventional models, our proposed model utilizes temperature and calendar information as primary inputs and can generate an electricity load without relying on the past load. This model’s design is rooted in the observation that temperature and calendar context sufficiently capture residential electricity consumption dynamics. By excluding the reliance on past load data, our model is able to synthesize electricity load for arbitrary temperature and calendar scenarios in addition to effectively predicting the future load. This synthesizing capability is invaluable for optimizing the planning and operation of residential energy systems, including photovoltaic systems and batteries. Experimental results using real-world electricity load data of residential buildings demonstrate the effectiveness of our transformer-based model, offering a robust framework for future load prediction and synthesis under varying conditions.

Keywords Electricity Load Forecasting, Electricity Load Synthesis, Transformer, Residential Electricity Load, Energy Management Systems