To optimize the heat sources layout in irregular three-dimensional integrated circuit (3D IC) for reducing the peak temperature, this paper proposes an improved particle swarm optimization (PSO) algorithm. Firstly, coordinates of heat sources are encoded as particle positions. Then the PSO algorithm is enhanced with dynamic inertia weight adjustment, elite selection, crossover, and mutation operators to iteratively refine the heat source layout. Deep neural network (DNN) is used to predict the 3D IC temperature distribution, which is trained using Sobol sampling for data generation, forward propagation, loss evaluation, and backpropagation. Its output is utilized as the fitness function in the improved PSO algorithm to inform the search for optimal heat source locations. Experimental results show that the proposed algorithm reduces the peak temperature of the four-layer 3D IC from 344.10 K to 325.71 K and that of the six-layer 3D IC from 355.01 K to 331.32 K. The DNN temperature prediction exhibits an error of less than 1 K compared to simulation results. Analysis of the optimized layouts indicates that large area heat sources significantly affect peak temperature, vertical overlaps lead to thermal accumulation requiring staggered placement.