In wireless sensor networks (WSNs), link is a key element to achieve interconnects and multi-hop communication. Link quality is the fundamental of upper protocols, such as topology control, routing, and mobile management. The effective link quality prediction (LQP) can not only improve networks throughput and decrease node energy consumption, but also prolong network life time. In this paper, we give a concrete analysis about the related works on WSNs link quality prediction. A novel model, fuzzy support vector regression (FSVR), is proposed to predict link quality, which makes the impact of noise and outliers get high accuracy. The link quality samples are collected from three different scenarios. Taking the character of data distribution in unstable links into consideration, a kernel fuzzy c-means (KFCM) algorithm as an unsupervised learning algorithm, is applied to cluster the training set automatically in terms of partition coefficient and exponential separation (PCAES). The membership degree of samples is obtained to get fuzzy set for FSVR. The chaos particle swarm optimization (CPSO) algorithm is employed on each cluster in order to choose the suitable parameter combination for the model. The experimental results show that compared with the empirical risk-based BP neural network prediction methods, the proposed prediction model achieves higher accuracy and better generalization ability.