Distributed sensors and interconnected processes give rise to intricate and high frequency event sequences in industrial monitoring systems. These events are critical for enabling proactive fault detection, maintenance scheduling, and operational optimization. Predictive reasoning facilitates these tasks. However, the prominent issue within industrial environments is capturing the intricate spatiotemporal dependencies owing to the limitations of RNNs and LSTMs. This paper presents a novel approach called Transformer-Based Temporal Graph Neural network (TGTN). Leveraging multi-head attention, the TGTN forms dynamic temporal graphs of event sequences and captures sensor and time node interdependencies. By imposing temporal encoding, graph construction, and transformer layers, the model learns contextual embeddings, significantly improving event prediction accuracy, and thus enhancing system interpretability. Empirical validation is performed using real world datasets from the industry which show the proposed model outperforms existing accuracy, robustness, and inference latency baselines. TGTN also demonstrates resilience to noisy signals, empty events, and complex topological structures. This study provides a robust framework for the exploration of deploying intelligent self-updating models for monitoring systems embedded within mission critical industries.