Accurate prediction of power usage is crucial for effective energy management, grid stability, and planning in smart grids. Traditional statistical methods often struggle to model the complex temporal dynamics and nonlinear demand patterns in energy consumption. To address these challenges, this work proposes a deep learning-based forecasting framework that utilizes state-of-the-art sequence modeling models for both short-term and long-term load prediction. Applied to a real dataset from a power plant in Maharashtra, which includes load demand, meteorological factors, and calendar features on an hourly basis over two years, the top three models evaluated are Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer. Data preprocessing involved normalization and a sliding window approach to train the models using the Adam optimizer. Model accuracy was assessed using MAE and RMSE. Results indicate that the Transformer model consistently outperforms recurrent models by capturing long-range dependencies. Additionally, supplementary analyses such as loss curves, error histograms, residual distributions, and horizon-based studies further demonstrate the robustness of the approach. In conclusion, this paper shows that the Transformer architecture is a valuable, scalable, accurate, and reliable tool for power consumption forecasting in smart grid applications.