A Data-Driven Framework for Grid Frequency Regulation Leveraging Aggregated Flexibility of Heterogeneous Resources Considering Network Delays

The rise of large-scale renewables has exacerbated frequency instability, revealing the limits of conventional frequency regulation frameworks in controlling deviations and incentivising fast-acting units (FAUs). Although the mileage-based payment framework promotes FAUs’ participation in automatic generation control (AGC) services, efficient instruction dispatch (ID) across an increasing number of heterogeneous AGC units remains challenging. The existing mileage-based payment framework lacks validation under intermittent generation or generation outages scenarios. This work proposes a data-driven ID framework with a modified payment scheme, adding a penalty term for refining mileage calculation to handle intermittent generation or generation outages during AGC operation. The framework uses a multihead attention-based encoder–decoder model, where the encoder extracts latent features and the decoder predicts unit-specific instructions. Attention mechanism improves accuracy by prioritising critical features, whereas L2 normalisation, dropout and k-fold cross-validation enhance models' robustness under unforeseen scenarios. The model aggregates the flexibility of multiple FAUs into a single entity, termed the FAU aggregator. Trained on a synthetic dataset generated from the evolutionary optimisation-based ID framework and validation on an interconnected system accounting for disturbance due to intermittent renewable energy sources' output, FAU variability and stochastic communication effects. The results demonstrate a reduction in both frequency deviation and area control error in comparison with other ID frameworks.