Dissolved gas analysis in transformer oil is regarded as an important indicator for evaluating the operational status of transformers. Accurate prediction of trends in dissolved gases in oil is beneficial for preventing power transformer failures. A Temporal Fusion Transformer (TFT) model, optimized via Optuna hyperparameter tuning, is proposed to address the technical challenge of low prediction efficiency inherent in traditional models that rely on a single variable. Static variables including transformer group, winding phase, and gas type are introduced into the model, and an interpretable multi-head attention mechanism is integrated as well. Synchronous prediction of all dissolved gases in the oil of multiple transformers is thereby achieved, improving the early warning efficiency of substation operation and maintenance systems. An average relative error of only 0.306% is achieved by the proposed model, representing a 66.7% reduction relative to the Transformer baseline model. Higher predictive accuracy is also demonstrated in both short-term and long-term forecasting. In addition, the model's training time is only one quarter that of the Transformer baseline model. This efficiency aligns with the current trend toward simultaneous prediction across multiple device groups in intelligent early-warning platforms. Strong correlations between hydrogen and methane and between carbon dioxide and methane are indicated by the model's multi-head attention mechanism. These correlations are consistent with the gas generation patterns of oil-paper insulation degradation, further demonstrating the model's good interpretability and providing technical support for synchronous prediction in multiple device groups.