Researchers from the University of Technology Sydney, including Fangzhou Lin, Guoshun He, Zhenyu Guo, Zhe Huang, and Jinsong Tao, have developed a new approach to improve electric load forecasting, which is crucial for efficient grid management and renewable energy integration. Their work, published in the journal Nature Energy, introduces a model called GRAFT (GRid-Aware Forecasting with Text) that leverages multiple sources of textual data to enhance forecasting accuracy.
Electric load, or the amount of power consumed, is influenced by various factors operating at different time scales. These include weather and calendar rhythms, sudden events, and policies. Traditional forecasting methods often struggle to incorporate these diverse influences effectively. GRAFT addresses this challenge by aligning daily-aggregated news, social media, and policy texts with half-hour load data. This alignment allows the model to guide the fusion of textual information to specific time positions using a mechanism called cross-attention, both during training and rolling forecasting.
One of the key features of GRAFT is its plug-and-play external-memory interface. This design allows the model to accommodate different information sources in real-world deployment, making it highly adaptable to various scenarios. The researchers constructed a unified benchmark covering the years 2019 to 2021 for five Australian states. This benchmark includes half-hour load data, daily-aligned weather and calendar variables, and three categories of external texts. The researchers conducted systematic evaluations at hourly, daily, and monthly scales under a unified protocol, ensuring comparability across regions, external sources, and time scales.
The experimental results demonstrated that GRAFT significantly outperforms strong baseline models and reaches or surpasses the state-of-the-art across multiple regions and forecasting horizons. Notably, the model is robust in event-driven scenarios, meaning it can handle sudden changes in load patterns effectively. Additionally, GRAFT enables temporal localization and source-level interpretation of text-to-load effects through attention read-out, providing valuable insights into how different textual sources influence electric load.
The researchers have released the benchmark, preprocessing scripts, and forecasting results to facilitate standardized empirical evaluation and reproducibility in power grid load forecasting. This work highlights the potential of integrating multi-source textual data into load forecasting models, offering practical applications for grid operators and energy providers. By improving forecasting accuracy, GRAFT can contribute to more efficient grid management, better integration of renewable energy sources, and enhanced overall energy system reliability.
This article is based on research available at arXiv.