Abstract: To address the measurement lag inherent in NOx online monitoring systems—where using the monitored selective catalytic reduction (SCR) inlet NOx concentration as the feedforward signal for ammonia injection leads to delayed injection control, which may cause ammonia slip or excessive NOx emissions—this paper proposes an SCR inlet NOx concentration prediction model based on a temporally aware dynamic causal graph convolutional network (TDCGCN) optimized by an improved particle swarm optimization (IPSO) algorithm, and further designs an SCR denitrification control strategy based on the prediction model. First, a simulated annealing algorithm is introduced into the conventional PSO to overcome its limitation of being prone to local optima, forming the IPSO algorithm. On this basis, the inlet NOx concentration prediction model is constructed using TDCGCN optimized by the IPSO algorithm. Second, the predicted inlet NOx concentration is incorporated as the ammonia injection feedforward signal into the control loop, establishing a PID plus feedforward denitrification control structure based on inlet NOx concentration prediction. Finally, simulation verification is carried out using real-time operational data from a 1000 MW coal-fired unit. The experimental results demonstrate that the TDCGCN network improves prediction performance by capturing the temporal and unidirectional causal characteristics of the SCR denitrification system, and that the proposed control strategy enables proactive adjustment of the ammonia injection rate, thereby enhancing the denitrification efficiency and stability of the outlet NOx concentration.