A hybrid framework for short-term load forecasting based on optimized InMetra Boost and BiLSTM

Abstract

Accurate short-term load forecasting is essential for maintaining stable and efficient power grid operations, especially with the increasing complexity introduced by renewable energy sources. This paper proposes a novel hybrid model for day-ahead Short-Term Load Forecasting, combining an improved Boosting algorithm, InMetra Boost, and a Bidirectional Long Short-Term Memory model. InMetra Boost introduces an asymmetric penalty mechanism, allowing for more precise handling of positive and negative forecast deviations. The model's hyperparameters are optimized using the Tree-structured Parzen Estimator, and the temporal dependencies in load data are further captured by a BiLSTM model, whose architecture is refined via the Cuckoo Search algorithm. The proposed TPE-IMB-CS-BiLSTM (Tree-structured Parzen Estimator optimized InMetra Boost within BiLSTM tuned by Cuckoo Search) framework was evaluated on real-world Estonian grid data, demonstrating superior performance compared to traditional models. Firstly, the TPE-IMB model achieves a significant improvement, reducing MAE from 34.86 MW in the original boosting model to 30.38 MW, and RMSE from 47.79 MW to 40.45 MW. Secondly, the hybrid TPE-IMB-CS-BiLSTM model further enhances accuracy, reducing MAE to 27.77 MW and RMSE to 36.55 MW, significantly outperforming the TPE-IMB and vanilla BiLSTM models. Lastly, compared to other state-of-the-art models, the proposed model achieves the best performance with a 24.66 % reduction in MAE and 26.74 % reduction in RMSE, demonstrating superior robustness in handling complex and extreme data conditions.