Fault prediction in network operations and maintenance (O&M) has become increasingly challenging due to the complexity and scale of modern AI workloads. Traditional topology-based AI for IT Operations methods fail to capture the temporal dynamics of fault propagation, which constrains accurate fault localization. Although knowledge graphs (KGs) have been introduced to alleviate this problem, they still fail to effectively capture static fault factual knowledge and neglect the temporal dynamics of network systems. To fill these gaps, this paper designs FFE-TKG, which fuses static fault factual knowledge with temporal information to build temporal KGs (TKGs). Based on the TKGs, a fault prediction architecture, FRP-TKG, is proposed, which leverages entity frequencies and contextual relationships in TKGs and employs reinforcement learning to enhance interpretability. Experiments on real datasets demonstrate that FRP-TKG outperforms existing baselines, enabling more effective and reliable O&M decision-making.