Improving Day-Ahead Grid Carbon Intensity Forecasting by Joint Modeling of Local-Temporal and Cross-Variable Dependencies Across Different Frequencies

TL;DR

This work tackles short-term grid CIF forecasting by framing it as a multivariate time-series problem and defining CIF with $CIF_{avg,t} = rac{ ext{\sum}_r (E_{r,t} \times C_r)}{ ext{\sum}_r E_{r,t}}$. It introduces two parallel modules: LT-MWKC, which captures local-temporal patterns across multiple frequencies using overlapping patches and diverse wavelet kernels, and CV-DWCC, which models dynamic cross-variable dependencies through wavelet-local multiple correlations and dominant-variable selection, integrated via 2D convolutions. A fusion stage with fully connected layers and a softmax weighting combines the modules for accurate day-ahead CIF forecasts. Experiments across four Australian markets show state-of-the-art performance and robust interpretability, with ablations confirming complementary benefits of LT-MWKC and CV-DWCC; a Grad-CAM case study demonstrates practical insights during atypical events, supporting real-world applicability for demand-side management and energy-scheduling decisions.

Abstract

Accurate forecasting of the grid carbon intensity factor (CIF) is critical for enabling demand-side management and reducing emissions in modern electricity systems. Leveraging multiple interrelated time series, CIF prediction is typically formulated as a multivariate time series forecasting problem. Despite advances in deep learning-based methods, it remains challenging to capture the fine-grained local-temporal dependencies, dynamic higher-order cross-variable dependencies, and complex multi-frequency patterns for CIF forecasting. To address these issues, we propose a novel model that integrates two parallel modules: 1) one enhances the extraction of local-temporal dependencies under multi-frequency by applying multiple wavelet-based convolutional kernels to overlapping patches of varying lengths; 2) the other captures dynamic cross-variable dependencies under multi-frequency to model how inter-variable relationships evolve across the time-frequency domain. Evaluations on four representative electricity markets from Australia, featuring varying levels of renewable penetration, demonstrate that the proposed method outperforms the state-of-the-art models. An ablation study further validates the complementary benefits of the two proposed modules. Designed with built-in interpretability, the proposed model also enables better understanding of its predictive behavior, as shown in a case study where it adaptively shifts attention to relevant variables and time intervals during a disruptive event.

Figures (7)

• Figure 1: An example from October 15, 2022, illustrates how building electricity demand can be shifted through a pre-cooling strategy during low levels of CIF periods (highlighted in green), reducing demand during high levels of CIF hours in the afternoon (highlighted in red). This helps align energy use with cleaner grid periods and contributes to emission reductions.
• Figure 2: Overall architecture of the proposed model.
• Figure 3: The architecture of LT-MWKC.
• Figure 4: The architecture of CV-DWCC.
• Figure 5: Forecasting results with MAE in four example cases: CIF with the largest variation in NSW (a) and SA (c); CIF with the smallest variation in NSW (b) and SA (d).