In the event of a severe nuclear accident, the measuring instruments inside the reactor may be damaged, posing significant challenges to directly obtaining the source term. However, rapid and accurate estimation of the source term is crucial to protect the public and the environment. Inversion methods based on backpropagation neural networks (BPNN) can effectively estimate source terms, but they are prone to falling into local minima. To address this, various intelligent optimization algorithms have been proposed to enhance the predictive performance of neural networks. This study compares the optimization performance of two recently proposed algorithms, the Crested Porcupine Optimization Algorithm (CPO) and the Dung Beetle Optimization Algorithm (DBO), on BPNN. Furthermore, it integrates the stronger global search capability of the Differential Evolution Algorithm (DE) with the local search capabilities of either the DBO or CPO algorithm. To further boost global search, a Tent chaotic mapping is applied to initialize the DE population. DE performs the global search, while DBO and CPO refine solutions through local search to further improve solution quality. This approach avoids local optima and fine-tunes solution precision. The methodology was validated in simulated release scenarios using meteorological and topographical data from the Sanmen Nuclear Power Plant. Results show that the Tent-DE-CPO-BP model achieved the best predictive performance, with a relative error of only 3.13% on test data. Therefore, this study recommends applying the Tent-DE-CPO-BP model to similar nuclear accident emergency response or consequence assessment scenarios.