A Fuzzy Prediction-based Dynamic Bandwidth Allocation (FPDBA) algorithm is proposed to enhance the differentiated services for EPONs based on the Prediction-based Fair Excessive Bandwidth Reallocation (PFEBR) in our previous work. The PFEBR proposed an Early-DBA mechanism which improves prediction accuracy by delaying report messages of unstable traffic ONUs and assign estimation credit to predict the traffic arrival during waiting time. However, delaying one report message will increase a guard time in one transmission cycle, how many report messages should be delayed and what is the optimal linear estimation credit are important issues. Both Fuzzy Unstable Degree List Controller (FUDLC) and Fuzzy Credit Estimator (FCE) mechanisms are incorporated to improve the prediction accuracy and enhance the system performance for differentiated services. The FUDLC chooses the second traffic variance and the mean traffic variance of ONUs as input linguistic variables to determine the optimal number of ONUs in the unstable degree list. In addition, the FCE chooses the degree of traffic variance and the degree of waiting time among ONUs as input linguistic variables for the credit estimation, so that the request bandwidth for the next cycle can be predicted more precisely. Simulation results show that the proposed FPDBA algorithm outperforms the efficient bandwidth allocation algorithm (EAA) and DBA with multiple services algorithm (DBAM) in terms of wasted bandwidth, gain ratio of bandwidth, throughput, downlink available bandwidth, average end-to-end delay and average queue length, especial in heavy traffic load.