Day-Type Segmented Load Forecasting with Multi-Head Attention-LSTM for Vehicle-to-Grid Scheduling

Accurate short-term load forecasting is critical for effective Vehicle-to-Grid (V2G) scheduling. This study proposes a day-type segmented forecasting approach, where two Multi-Head Attention Long Short-Term Memory (MHA-LSTM) models are independently trained on weekday and weekend data. This design is motivated by the clear differences in load patterns across day types, such as regular weekday peaks driven by commuting routines versus the more erratic weekend profiles. Modeling them jointly may obscure such trends and degrade scheduling outcomes. To further enhance trend recognition, a Load Relative Magnitude (LRM) feature is introduced, allowing the model to focus on deviations from daily averages. Results demonstrate that the segmented models consistently yield lower V2G-SVE scores, validating the benefit of day-type-specific modeling in improving practical V2G scheduling performance.