Isolated power grids in new green electrolytic aluminum industrial parks are characterized by large loads, high proportion of renewable energy, and limited regulating capability. Their operation is easily affected by uncertainties such as sudden weather changes and anode effects, resulting in severe fluctuations in photovoltaic output and electrolytic aluminum load that threaten grid security. To address these issues, this paper proposes a two-stage multi-timescale optimal scheduling method consisting of day-ahead and intra-day phases. In the day-ahead phase, the objective is to minimize system operating cost by optimizing thermal power and storage output plans. In the intra-day phase, both economic efficiency and minimal adjustment effort are considered. A rolling optimization strategy based on model predictive control (MPC) is adopted to update the schedule, while leveraging the virtual battery characteristics of electrolytic aluminum load to mitigate source-load power fluctuations and forecasting errors. To overcome the limitation of conventional fixed-step MPC, which struggles to balance optimization accuracy and computational efficiency, an event-triggered adaptive variable-step MPC strategy is proposed. This strategy dynamically adjusts the scheduling step size according to real-time power fluctuations and prediction errors. Case studies show that the proposed method effectively reduces photovoltaic curtailment, alleviates thermal unit regulation pressure, and enhances the flexibility, economy, and sustainability of electrolytic aluminum industrial park operations.