With the goal to realize the real-time accurate measurement of chemical oxygen demand ( COD) in wastewater treatment process, a soft-measurement method based on TentFWA-GD RBF neural network (NN) was proposed. To solve the problems of network parameters settings and local optima existing in RBF NN based soft sensor modeling for complex industrial processes, as well as improve the model’s prediction precision and generalization ability, tent chaotic mapping was introduced in fireworks algorithm (FWA) to keep the population diversity and avoid the premature convergence by making use of the global ergodicity of chaos movement. Then a novel training method for RBF NN was proposed by combining the improved TentFWA with gradient descent (GD) method to enhance the learning ability. The TentFWA-GD RBF NN was applied to construct the fitting models of four Benchmark functions and the COD soft sensor model of rural domestic sewage treatment process. Simulation and application results showed that the model had lower function approximate error and higher COD prediction precision as compared with other neural network models. In COD soft sensor modeling, the mean square error and mean absolute error of the training results were 0. 18 and 0. 25, which of the test results were 0. 23 and 0. 36, respectively.