To address the practical challenges of poor data quality, strong hyperparameter coupling, and significant peak positioning errors in power system carbon emission peak forecasting, a framework integrating robust data preprocessing with an improved grey-convolution hybrid model is proposed. Firstly, a dynamic quantile boundary-based outlier detection and multi-window weighted robust repair procedure is established, together with a random forest feature importance-based chained multiple imputation method, to suppress outlier disturbances and high-dimensional missingness in non-Gaussian data. Subsequently, an improved convolutional network incorporating variational mode decomposition, dilated convolution, and attention gating is constructed, with an embedded improved grey model for long-term trend extraction; hyperparameter optimization is achieved through a grey relational grade-guided whale optimization algorithm. Experimental results show that compared with genetic algorithm, particle swarm optimization, and grey wolf optimization, the proposed algorithm reduces mean absolute percentage error by 39.7%, 32.5%, and 25.4%, and peak prediction time deviation by 77.1%, 71.4%, and 60.0%, respectively. Compared with autoregressive integrated moving average (ARIMA), long short-term memory - prophet (LSTM-Prophet), time-series transformer (TST), empirical mode decomposition-LSTM (EMDE-LSTM), and variational mode decomposition - gated recurrent unit (VMD-GRU), the proposed model reduces mean absolute percentage error to 2.89% and peak prediction time deviation to 0.7 h, while improving the inflection point capture rate to 93.8%. This study provides a new technical approach for accurate carbon emission peak prediction and offers data support for power system emission reduction strategy formulation.