GridNet: A Hybrid LSTM-Based Deep Learning Approach for Accurate Electricity Consumption Forecasting in Smart Grid Systems

Abstract

Accurate forecasting of power production and consumption is crucial for optimizing smart grid operations, especially with the growing integration of renewable energy sources, and minimizing CO₂ emissions. This study develops GridNet, a deep learning-based model for forecasting next-hour electricity consumption using the past 24 hours of electricity and time features. GridNet utilizes a two-layer stacked Long Short-Term Memory (LSTM) network with dropout layers to prevent overfitting and improve generalization, followed by a dense output layer. Early stopping and learning rate reduction callbacks were applied to enhance convergence efficiency. The model was optimized using the Adam optimizer and a hybrid loss function combining Mean Squared Error (MSE) with L1 regularization for improved prediction accuracy and robustness. GridNet was trained on a six-year hourly dataset from Romania, covering diverse energy sources like nuclear, wind, solar, for 30 epochs. The performance of the proposed model GridNet was evaluated using several standard metrics, including the Coefficient of Determination (R2), Mean Absolute Error (MAE), MSE, and Root Mean Squared Error (RMSE). Additionally, GridNet's effectiveness was compared with state-of-the-art algorithms, including XGBoost, KNN, Random Forest, and SVM. GridNet outperforms all competing algorithms by up to 59.9% in R2, 72.0% in MAE, 89.8% in MSE, and 68.1% in RMSE, demonstrating its superior accuracy in forecasting electricity consumption trends in smart grids.