Enhancing smart grid load forecasting: An attention-based deep learning model integrated with federated learning and XAI for security and interpretability

Abstract

• Developed an optimized attention-based 1D-CNN-GRU model for secure and effective load forecasting in smart grid systems, employing federated learning to enhance data privacy and security in collaborative environments. • Improved model accuracy and efficiency through hyperparameter optimization using particle swarm optimization (PSO) and enhanced training and generalization with data preprocessing and augmentation techniques. • Enhanced model interpretability with SHapley Additive exPlanations (SHAP), offering insights into factors affecting load predictions, and analyzed the impact of various factors through feature ranking to inform stakeholders in the energy sector. • Modified the federated learning aggregation mechanism with pruning-based methods, reducing model parameters and computational costs while maintaining performance. • Outlined potential future research directions to address existing limitations and further improve the proposed framework for load forecasting in smart grids. Smart grid is a transformative advancement that modernized the traditional power system for effective electricity management, and involves optimized energy distribution by load forecasting. Precise load forecasting provides the best utilization of energy resources and increases sustainability. Dynamic changes of several connected factors, such as temporal and geographical variability, pose challenges to accurate load prediction. Integrating Artificial Intelligence (AI) in the smart grid can enhance the performance of the forecasting process by capturing these changes. This study investigated load forecasting tasks on four different datasets. Several preprocessing and augmentation techniques are applied to increase the data quality. An attention-based 1D-CNN-GRU model is proposed to capture the temporal patterns from the time-series data, and the hyperparameters of the model are optimized using a particle swarm optimization (PSO) algorithm that also accelerates the convergence and results in an efficient training session. Empirical evaluations highlight that the proposed model substantially minimizes the loss, reflecting the ability to make accurate predictions. It obtains MAE values of 0.12, 0.8, 16.48, and 82.64 for the four datasets. Moreover, the explainable AI (XAI) technique is applied using Shapley Additive explanations (SHAP) to interpret the model prediction, providing the feature ranking based on their prediction score. Moreover, this study utilizes federated learning, enables collaborative training, maintains the privacy of the grid data, and secures the process comprehensively. The aggregation mechanism in federated learning is modified using pruning-based methods that reduce the parameters and computational cost, resulting in a more efficient framework. Integrating all these approaches provides valuable insights for developing a load forecasting model and outlines potential contributions in the smart grid domain.

Topics & Concepts