Effective energy forecasting is essential for reducing challenges related to smart grid systems. In this paper, a bidirectional long short-term memory network complemented with an attention mechanism is used to propose an advanced method of energy consumption prediction. We built a model that can capture both sequential dependencies and the relative relevance of distinct time steps, using data from the Victorian energy dataset and the AEP dataset. To increase prediction accuracy and generalization, the architecture combines a preprocessing stage followed by several LSTM layers, batch normalization, dropout layers, and Luong attention mechanisms. When attention mechanisms are taken into account, evaluation metrics like Mean Absolute Error and Root Mean Squared Error show considerable gains in predicting performance. The proposed method surpasses all state-of-the-art approaches, achieving an RMSE of 0.0189 and MAE of 0.0235 for the Victorian energy consumption dataset and an RMSE of 0.0129 and MAE of 0.0117 on the AEP dataset. Based on our research, attention mechanisms and LSTM networks can be used to improve energy consumption prediction, offering a powerful tool for energy forecasting and management.