Improving long-term electricity time series forecasting in smart grid with a three-stage channel-temporal approach

Abstract

Listen

Translate article

—Transformer-based models have shown progress in addressing electricity time series forecasting challenges. However, as the forecasting horizon extends, the computational complexity required to capture long-term global correlations may limit their ability to utilize extensive historical data. This paper proposes a non-Transformer model named Three-Stage Channel-Temporal (TSCT), designed to be lightweight and capable of handling longer look-back windows for long-term electricity time series forecasting (LTESF) in smart grid contexts. TSCT sequentially derives feature maps along two dimensions, channel and temporal, focusing on ‘which' and ‘when', respectively. Moreover, its dynamic capacity to decompose and fuse information enables the disentanglement of intricate temporal patterns, highlighting the fundamental characteristics inherent in the time series. Extensive experiments demonstrate that our proposed TSCT outperforms state-of-the-art methods in smart grid scenarios using a commonly used Electricity dataset. Notably, the TSCT approach exhibits significantly higher efficiency compared to Transformer-based methods: an impressive 85% reduction in trainable parameters, a substantial 99% reduction in GPU memory usage, a 94% reduction in running time, and a 49% reduction in inference time. Code is available at: https://github.com/Zhao-Sun/TSCT.