C-NET: Enhanced Temporal Residual Convolution for Carbon Price Forecasting with Explainable Interpretability

The carbon price fluctuation is based on market dynamics and regulatory constraints, directly influencing the behaviour of the power sector. This fluctuation often makes enterprises and companies make incorrect decisions on profits. This study introduces a novel deep learning model, the Carbon Network (C-Net), specifically the Temporal Residual Harmonics Network (TRHN), for forecasting carbon prices using multi-source monthly official time-series data collected from January 2015 to July 2025. Data were aggregated from publicly available resources, including Ember, which covers electricity generation, demand, emission intensity, and sectoral CO2 emissions. Feature Engineering is computed using lag computation, rolling mean and standard deviation features. The TRHN architecture integrates sinusoidal temporal harmonics, dilated residual convolutions, and shock gating mechanisms to capture both periodic trends and long-range dependencies in the data. The model is trained using 5-fold cross-validation and evaluated with MAE (2.71), RMSE (3.18), and $R^2$ (0.9916) metrics. It successfully forecasts carbon prices for the second half of 2025, demonstrating strong predictive performance and reliability. This work highlights the potential of deep temporal models in carbon market forecasting and policy support across the EU power sector.