Water temperature is considered to be the most important parameter which can largely determine the aquaculture production of sea cucumbers, so it is extremely important to monitor and forecast the water temperature at different water depths. As the change of water temperature is a complex process which can not be exactly described with a certain formula, the artificial neural network characterized by non-linearity, adaptivity, generalization, and model independence is a proper choice. This paper presents a RBF neural network model based on nearest neighbor clustering algorithm and puts forward four improved methods, then integrates them into an optimization model and verifies it on matlab platform. Finally, a comparison between the optimized RBF model and the original RBF model is made to confirm the excellent forecasting performance of the optimized RBF neural network model. This paper provides a relatively impeccable learning algorithm to complete the choice of radial basis clustering center in the process of RBF network design, and obtains a high forecasting precision so that the demand of water temperature forecasting in sea cucumber aquaculture ponds can be satisfied.