Disruption in a tokamak nuclear reactor refers to the rapid extinction of the plasma confinement. This is often an uncontrolled event that involves the loss of plasma stability and can potentially cause damage to the reactor itself. To ensure the safety of fusion reactors, precise disruption prediction for early identification is crucial. While numerous data-driven time-series models have been developed and are continuously evolving to enhance disruption prediction in tokamaks, these models however often rely on fixed time windows for predictions. Because of the dynamic nature of plasma discharge, traditional models like LSTM, Bi-LSTM, and Stacked LSTM often produce premature alarms that make forecasts too early to determine if a signal reliably indicates a disruption. In this study, we propose a novel dynamic time window aggregation mechanism integrated with a sequential Bi-LSTM model (Bi-LSTM-DTWA), for predicting disruptions. By dynamically adapting to each signal time, this approach enhances prediction performance and effectively addresses the issue of premature alarms. The implemented model is trained using data from the medium-sized Aditya tokamak. Experimental validation on the Aditya dataset, comprising 153 disruptive shots and 67 normally terminated shots with nine diagnostic signals each, shows that the predictive model efficiently forecasts disruptions within 10 to 23 ms in advance without premature alarms, making it suitable for real-time deployment with minimal computational overhead.