Aiming at the shortcomings of existing water quality prediction models in data noise reduction, initial setting and optimization of network parameters, and accuracy improvement, an optimized three-dimensional water quality prediction model was constructed. The key parameters of water quality were screened by using the principal component analysis algorithm, the three-dimensional water quality parameters and meteorological data were de-noised based on the fully set empirical mode decomposition algorithm based on adaptive noise combined with the wavelet threshold model, the feature data set was extracted by using the three-dimensional convolutional neural network (3D CNN) and the dynamic initial values of hyperparameters in radial basis function (RBF) neural networks were optimized by an improved cuckoo search algorithm (ICS) based on autoencoder (AE). The comparison and verification results of the measured data from twenty-two typical online monitoring stations and six handheld monitoring stations in the Dashuiqiao Reservoir area of Xuwen County, Zhanjiang City, Guangdong Province showed that turbidity and algae density were positively correlated with total nitrogen, and chlorophyll was positively correlated with temperature, the proposed water quality prediction model was superior to the existing literature methods in five typical accuracy evaluation indicators. The research results can provide reference for management departments and researchers to monitor water quality. Introducing inertial weight and adjusting position parameters to improve CS to speed up the convergence of RBF network. The autoencoder was used to initialize the initial values of network parameters to avoid the defects of artificial random setting. Adding WT algorithm can effectively reduce the white noise in the decomposition and reconstruction process of the fully set empirical mode decomposition algorithm based on adaptive noise.