Abstract: Power carbon emission intensity is an important indicator for measuring the carbon emission level of the power system. Accurate calculation and prediction of power carbon emission intensity are of great significance for effectively implementing carbon emission management strategies and scientifically formulating emission reduction plans. To this end, a dynamic calculation and prediction method for carbon emission intensity is proposed. Firstly, based on the functional relationship between coal consumption and load of coal-fired units, a dynamic carbon emission intensity model for coal-fired units is constructed. Then, by integrating the output characteristics of power generation units and the dynamic carbon emission intensity, a real-time calculation model for grid-level carbon emission intensity is established. On this basis, convolutional neural networks (CNN) are used to extract data features of dynamic carbon emission intensity, which are then input into a bidirectional long short-term memory network (BiLSTM). Additionally, an improved sparrow optimization algorithm (ISSA) is introduced, integrating strategies including circle population initialization, adaptive factors, Cauchy mutation, and Sine mapping perturbation to optimize the hyperparameter configuration of BiLSTM. Finally, a hybrid model of ISSA-CNN-BiLSTM is constructed to achieve high-precision short-term prediction of carbon emission intensity. Simulation and test results show that compared with models such as long short-term memory networks (LSTM), the proposed ISSA-CNN-BiLSTM model exhibits higher prediction accuracy and stronger generalization capabilities in carbon emission intensity prediction.