Hybrid CNN–LSTM Model With Soft Attention Mechanism for Short-Term Load Forecasting in Smart Grid

Abstract

Integrating renewable energy in smart grids enables sustainable energy development but introduces challenges in supply–demand variability. Deep learning techniques are now imperative for Short-Term Load Forecasting (STLF), a significant enabler of energy flow management, demand-side flexibility, and grid stability. These methods optimize smart grid performance under variable conditions by leveraging the synergistic integration of multiple architectures. This paper proposes a novel hybrid CNN–LSTM parallel model with a soft attention mechanism to improve smart grids' STLF. The proposed model leverages Convolution Neural Networks (CNNs) to extract spatial patterns, LSTMs to capture temporal dependencies, and attention mechanisms to prioritize important information, enhancing predictive performance. A comprehensive comparative analysis uses two publicly available datasets, American Electric Power (AEP) and ISO New England (ISONE), to evaluate the proposed model's effectiveness. The proposed model provides outstanding performance across single-step and multistep forecasting operations by delivering the highest evaluation results. The proposed model delivered single-step forecasting results of 123.91 Root Mean Square Error (RMSE), 92.8 Mean Absolute Error (MAE), and 0.63 Mean Absolute Percentage Error (MAPE) on the AEP dataset and 126.16 RMSE, 64.28 MAE, and 0.44 MAPE on the ISONE dataset. The model delivered multistep forecasting results on AEP, which showed RMSE at 685.25, MAE of 490.37, and MAPE of 3.27, while ISONE produced RMSE of 598.26, MAE of 402.44, and MAPE of 2.73. The simulation results demonstrate that parallel CNN–LSTM with a soft attention mechanism effectively supports the development of adaptive and resilient smart grids, enabling better integration of renewable energy sources.