Event-Driven CNN–LSTM for Grid Exchange Forecasting in Industrial Microgrids: Feature Ablation and Multi-Horizon Evaluation

Industrial microgrids integrating distributed photovoltaic (PV) systems require accurate grid exchange forecasting for effective energy management. However, grid exchange exhibits substantial fluctuations due to physical events such as cloud shading, PV curtailment, and surplus export. Most existing deep learning methods rely primarily on raw historical measurements and overlook the physical causes of these variations. This study proposes an event-driven CNN-LSTM framework that augments raw PV and grid signals with a PV ramp rate feature to capture abrupt transitions in solar output. Seven operational indicators are extracted from measured microgrid data; however, an empirically selected three-feature input set comprising PV generation, grid exchange, and PV ramp rate is used for aggregate forecasting, while the remaining event flags are assessed separately through feature ablation and event period analysis. The framework is evaluated using one year of hourly day-ahead data and one month of 15 min very-short-term data from a 998.58 kWp PV installation. Compared with baseline models, the proposed model achieves the best DA forecasting performance, with an RMSE of 118.26 kW and R2 of 0.8463. Ablation analysis confirms that raw PV generation provides the dominant predictive information, whereas the ramp rate feature yields targeted improvements during PV disturbances. Event period analysis further shows that binary event flags provide useful information during specific physical events but may reduce aggregate forecasting accuracy when permanently incorporated into the model.