Predicting load profiles is crucial for efficient electricity management, but traditional methods often struggle with the complexity of real-world data. While deep learning models have been explored for load forecasting, achieving high accuracy remains challenging. This paper presents a streamlined ensemble approach that combines bidirectional LSTM (BiLSTM), bidirectional GRU (BiGRU), and temporal convolutional network (TCN) layers to capture intricate temporal patterns in load profiles. A self-attention mechanism enhances the model's focus on the most relevant features, improving overall representation. The outputs of these components are combined using an XGBoost regressor to produce the final prediction. In testing, this hybrid model achieved notably higher accuracy up to 93% and faster processing times than other advanced models, showing strong promise for real-time load forecasting in smart grid systems.