[Objective

]

Coal-fired power plants play a pivotal role in China's power system

and accurate forecasting of their CO

2

emissions is crucial for achieving energy-saving and emission-reduction targets. To address this challenge

this paper proposes a short-term CO

2

emission prediction model for coal-fired power plants based on CNN-BiLSTM-Attention. [Methods

]

First

feature selection is performed by analyzing the correlation between CO

2

emission rates and various input variables

including unit load

SO

2

emission rate

NO

x

emission rate

and environmental factors

thereby reducing redundant information that may impair model training. Specifically

the Pearson correlation is used to quantify the linear relationship between each candidate feature and the CO

2

emission rate

and features with a correlation coefficient below 0.2 are excluded. The raw operational data are then normalized to the range [0

1

]

and segmented into samples using a sliding time window to capture temporal dynamics. Subsequently

the selected features are fed into the proposed CNN-BiLSTM-Attention model. In this hybrid architecture

the CNN layer extracts local patterns from the input sequence

the BiLSTM layer captures both forward and backward temporal dependencies

and the attention mechanism assigns adaptive weights to the most informative time steps. To further enhance performance

an improved Kepler optimization algorithm(IKOA) is introduced to optimize the model's hyperparameters. The model is validated using real-world operational data from coal-fired power plant units. In addition to root mean square error(RMSE)

the mean absolute error(MAE) and the coefficient of determination(R

2

) are also computed to comprehensively evaluate prediction performance. Comparative experiments show that among LSTM

BiLSTM

CNN-BiLSTM

and CNN-BiLSTM-Attention

the last model achieves the lowest RMSE of 24.329 2 t/h

compared to 26.261 0 t/h

26.039 3 t/h

and 25.429 3 t/h for the other three models

respectively. Moreover

when optimizing the CNN-BiLSTM-Attention model with four metaheuristic algorithms(genetic algorithm(GA)

whale optimization algorithm(WOA)

Kepler optimization algorithm(KOA)

and IKOA)

the IKOA yields the best fitness value of 21.584 1 t/h

slightly outperforming WOA(21.594 6 t/h)

KOA(22.136 0 t/h)

and GA(22.807 2 t/h). Results demonstrate that the proposed model exhibits strong transferability across units of similar capacity and effectively captures CO

2

emission trends over time

although prediction errors increase during seasonal transitions due to variations in operating conditions and coal quality.[Results

]

The proposed model consistently outperforms baseline models in both prediction accuracy and stability.[Conclusion

]

The CNN-BiLSTM-Attention model enables effective CO

2

emission forecasting for coalfired power plants even without direct CO

2

measurements

showing significant applicability and potential for broader deployment. This capability is particularly valuable for plants lacking continuous emission monitoring systems(CEMS)

offering a cost-effective alternative for CO

2

estimation. However

long-term or cross-seasonal predictions require further consideration of external factors such as coal quality. This study is primarily validated on limited units from a single plant; thus

its generalizability across different regions

coal sources

and operational strategies warrants further investigation. Future work will incorporate direct coal-quality indicators through multisource data fusion or online updating mechanisms to enhance long-term predictive performance and robustness.