Advanced Feature Engineering and Ensemble Methods for Lambda Microgrid Load Forecasting with Bidirectional Power Flow

In this study, a machine learning-based microgrid load forecasting framework is proposed to solve the problem of bidirectional power flow and high renewable penetration. A sophisticated feature engineering pipeline is developed by using feature transformations such as power direction, cyclic time conversion, degree-hours and interaction terms. This proposed approach is evaluated and compared with five machine learning models such as Random Forest, XGBoost, LightGBM, Gradient Boosting as well as an Ensemble. The results obtained on real microgrid data showed strong performance. The Random Forest model achieved the R2 score of 0.7109 and the mean absolute error of 0.7462 kW. The ensemble model is highly robust and obtained the R2 score of 0.7081. These results indicate that the proposed method is very beneficial for practical use in the real microgrid operation and control design. This study presents a practical feature engineering approach and designs a new performance baseline for a microgrid load prediction. Instead of proposing a new model architecture, the key novelty of this research is a domain-informed feature engineering framework for bidirectional microgrid power flows and a systematic benchmarking of ensemble models on real working microgrid data.