The growing penetration of variable renewable energy sources (vRES) and electrification-driven demand is increasing the volatility of real-time system operation, highlighting the need for early-warning tools to anticipate periods of stressed balancing. This work proposes a data-driven framework to forecast upward and downward manual frequency restoration reserve (mFRR) activation events in the Portuguese power system using five alert levels, from Level 1 to Level 5 (most critical), defined from historical data to provide a progressive and operationally meaningful characterisation of reserve stress. mFRR activation is predicted using a machine-learning classifier.

Results indicate that wind power forecast deviations are the dominant driver of mFRR activation. Detection precision increases from nearly 50% for the most critical events (Level 5) to nearly 80% for less critical levels. Upward activation is strongly dependent on forecast horizon, with better performance closer to real-time operation, while downward activation shows similar performance for all horizons analysed. Overall, the framework can complement existing deterministic forecasting and operational tools by enabling earlier assessment of system conditions and supporting adaptive reserve allocation in increasingly vRES-dominated systems.