Accurate forecasting of carbon prices is crucial for policymakers and market participants, yet remains challenging due to volatility, non-stationarity, and structural breaks in carbon markets. Data preprocessing plays a key role in improving predictive accuracy, but conventional normalization methods such as z-score and min–max rely on static statistics and fail to adapt to evolving dis- tributions. This study introduces Extended Deep Adaptive Input Normalization (EDAIN) into carbon price forecasting. EDAIN dynamically integrates outlier suppression, shifting, scaling, and power transformation, optimized jointly with forecasting models. We evaluate EDAIN with GRU, LSTM, and XGBoost on datasets from KRBN and KCCA obtained via the yfinance API. Experimental results show that EDAIN consistently outperforms static normalization, reducing the RMSE from 2.45 to 0.80, the MAE from 2.09 to 0.63, and the MAPE from 7.07% to 2.13% (KRBN dataset), with similar improvements observed on KCCA (RMSE reduced from 2.78 to 0.88, MAE from 2.30 to 0.67, MAPE from 7.48% to 2.22%). These findings highlight EDAIN’s methodological contribution to adaptive preprocessing and its practical value for improving risk management and decision-making in carbon trading.