Periodicity-Aware Time Series Forecasting for Electric Load

Accurate electricity load forecasting plays a pivotal role in power system scheduling and energy management, as it is critical for enhancing the operational efficiency of power systems and ensuring the security of electricity supply. This paper proposes a long-term time-series forecasting method for electricity load prediction, named DETformer, which focuses on optimizing the modeling of periodic characteristics. Electricity load time-series data typically exhibit prominent main cycles, such as daily cycles. Therefore, this paper employs the Periodicity Extraction Module (PEM) to explicitly model the most significant periodic patterns. Since the backbone network alone cannot accurately capture long-period cycles, we introduce the Timestamp Feature Enhancement Module (TFEM) to supplement temporal feature information, thus improving the model’s ability to predict special periods in long cycles, such as weekends and holidays. Additionally, given that periodic signals exhibit concentrated energy in specific frequency components in the frequency domain, we incorporate the Frequency Domain Constraint Method (FDCM) into the loss function to force the model to learn the periodic characteristics. Experimental results show that DETformer outperforms several benchmark models in terms of prediction precision and robustness, demonstrating its practical applicability in real-world power systems.