Short-term CO2 emissions forecasting: insight from the Italian electricity market

This study investigates the short-term forecasting of carbon emissions from electricity generation in the Italian power market. Using hourly data from 2021 to 2023, several statistical models and forecast combination methods are evaluated and compared at the national and zonal levels. Four main model classes are considered: (i) linear parametric models, such as seasonal autoregressive integrated moving average and its exogenous variable extension; (ii) functional parametric models, including seasonal functional autoregressive models, with and without exogenous variables; (iii) (semi) non-parametric and possibly non-linear models, notably the generalised additive model (GAM) and TBATS (trigonometric seasonality, Box-Cox transformation, ARMA errors, trend, and seasonality); and (iv) a semi-functional approach based on the K-nearest neighbours. Forecast combinations include simple averaging, the optimal Bates and Granger weighting scheme, and a selection-based strategy that chooses the best model for each hour. The results show that the GAM produces the most accurate forecasts during the daytime hours, while the functional parametric models perform best in the early morning. Among the combination methods, the simple average and the selection-based approaches consistently outperform all individual models. The findings underscore the value of hybrid forecasting frameworks in improving the accuracy and reliability of short-term carbon emissions predictions in power systems. In addition, they highlight the importance of considering zonal specificities when implementing flexible energy demand strategies, as the timing of low-carbon emissions varies between market zones throughout the day.

💡 Research Summary

This paper investigates one‑day‑ahead forecasting of carbon dioxide (CO₂) emissions from electricity generation in the Italian power market using hourly data spanning 2021‑2023. The authors analyse both the national aggregate and seven market zones (North, Centre‑North, Centre‑South, South, Calabria, Sicily, Sardinia) to capture regional differences in generation mix. Emissions are computed from fuel‑specific generation data supplied by ENTSO‑E, combined with Italian emission factors and oxidation rates. Preliminary statistical checks (two‑week rolling means, standard deviations, first‑differences) reveal no structural breaks, confirming stationarity of the differenced series.

Four families of statistical models are evaluated: (i) linear parametric models – seasonal ARIMA with exogenous variables (SARIMAX); (ii) functional parametric models – seasonal functional autoregressive (SFAR) and its exogenous‑variable extension (SFARX); (iii) (semi) non‑parametric and possibly non‑linear models – generalized additive model (GAM) and TBATS (trigonometric seasonality, Box‑Cox transformation, ARMA errors, trend, and seasonality); and (iv) a semi‑functional approach based on K‑nearest neighbours (KNN). All models are calibrated on a rolling training window and produce 24‑hour forecasts for the next day.

Forecast performance is assessed with hourly average root‑mean‑square error (RMSE), out‑of‑sample R², Diebold‑Mariano (DM) tests, and the Model Confidence Set (MCS) procedure. Results show a clear time‑of‑day pattern: GAM delivers the lowest RMSE and highest R² during daytime hours (approximately 08:00‑20:00), reflecting its ability to capture non‑linear relationships between demand, renewable output, and weather. In contrast, the functional models (SFAR, SFARX) excel in the early‑morning window (00:00‑06:00) where the emission profile follows a smoother, more regular seasonal shape. TBATS and KNN provide intermediate performance; TBATS is particularly useful on weekends and holidays where multiple seasonal cycles interact. SARIMAX, while robust, is outperformed by the more flexible models in most periods.

Three forecast combination schemes are examined: (1) simple averaging of all individual forecasts, (2) optimal weighting following Bates and Granger (1969), and (3) a selection‑based approach that picks, for each hour, the model with the smallest validation error. The simple average and the selection‑based method consistently outperform any single model, with the selection‑based scheme achieving a 5‑7 % reduction in overall RMSE relative to the best individual model. The Bates‑Granger weighting, despite its theoretical optimality, yields only marginal gains due to estimation instability in this context.

Regional analysis reveals that the optimal model varies across zones because of differing generation mixes. Zones with higher renewable penetration (South, Sardinia) see GAM perform better even in early‑morning hours, while zones dominated by thermal generation (North, Centre‑North) benefit more from functional models during low‑demand periods. These findings underline the importance of zone‑specific forecasting for demand‑side flexibility programs such as smart electric‑vehicle charging or storage dispatch.

The study contributes to the literature by (a) providing the most extensive comparison of linear, functional, semi‑non‑parametric, and nearest‑neighbour models for short‑term CO₂ emission forecasting in a European market, (b) demonstrating that straightforward combination techniques can surpass sophisticated weighting schemes, and (c) highlighting the practical relevance of zonal forecasts for emission‑aware demand‑response strategies. The authors suggest future work to integrate real‑time renewable forecasts and electricity price signals into a multivariate hybrid framework, and to translate improved emission forecasts into optimal operational decisions for flexible loads and storage assets.