Adaptive filter-driven optimized attention-based CNN-LSTM for load forecasting in microgrids
Stéfano Frizzo Stefenon, Kin‐Choong Yow
Abstract
Load forecasting in microgrids enables efficient balancing of supply and demand, ensuring that energy generation, storage, and consumption are optimally coordinated. A major challenge in load forecasting in microgrids is that the optimal model for one grid may not be the best for another, considering each system's different characteristics. Based on this concern, this paper proposes a data-driven adaptive filter, ensuring that the model can be applied to any microgrid. For model tuning, the adaptive tree-structured Parzen estimator (ATPE) was shown to be more efficient in finding the optimal hyperparameters than random search, annealing search, and TPE optimization strategies. The proposed hybrid prediction method integrates an adaptive filter (AF) input stage into an optimized attention-based (OA) convolutional neural network with long short-term memory (CNN-LSTM). Based on that, the model features a data-driven AF that automatically adjusts its denoising hyperparameter based on the input signal's sampling rate, ensuring robust performance across diverse datasets without manual tuning. When evaluated on three microgrid datasets (Liege, Technical University of Ostrava, and Rye), the proposed AF-OA-CNN-LSTM model demonstrated top performance compared to state-of-the-art deep learning architectures. Achieving an RMSE of 0.00116 (3.44% better than DeepAR and 239.65% better than TimesNet, the 2nd and 3rd best models, respectively) and a MAPE of 1.15% (296.52% better than TFT and 313.04% better than TimesNet, the 2nd and 3rd best models, respectively) in the best case (Liege dataset), the proposed method is a promising generalizable solution for load forecasting in different load contexts. • Propose an adaptive filtered attention-based CNN-LSTM model optimized via ATPE. • Developed a data-driven adaptive filter generalizable for load forecasting. • APTE for hypertuning compared to random search, annealing search, and TPE. • The hybrid approach achieved superior performance across three load datasets.