Experimental Comparison of Two Main Paradigms for Day-Ahead Average Carbon Intensity Forecasting in Power Grids: A Case Study in Australia

Accurate carbon intensity forecasts enable consumers to adjust their electricity use, reducing it during high fossil-fuel generation and increasing it when renewables dominate. Existing methods for carbon intensity forecasting can be categorized into a source-disaggregated approach (SDA), focused on delivering individual generation forecasts for each potential source (e.g., wind, brown-coal, etc.), and a source-aggregated approach (SAA), attempting to produce a single carbon intensity forecast for the entire system. This research aims to conduct a thorough comparison between SDA and SAA for carbon intensity forecasting, investigating the factors that contribute to variations in performance across two distinct real-world generation scenarios. By employing contemporary machine learning time-series forecasting models, and analyzing data from representative locations with varying fuel mixes and renewable penetration levels, this study provides insights into the key factors that differentiate the performance of each approach in a real-world setting. The results indicate the SAA proves to be more advantageous in scenarios involving increased renewable energy generation, with greater proportions and instances when renewable energy generation faces curtailment or atypical/peaking generation is brought online. While the SDA offers better model interpretability and outperforms in scenarios with increased niche energy generation types, in our experiments, it struggles to produce accurate forecasts when renewable outputs approach zero.