Accurate time series forecasting is essential in energy management, finance, supply chain optimization, and healthcare, where reliable predictions inform critical decision-making. While traditional models like ARIMA and LSTMs have laid the groundwork, they often fail to capture long-range dependencies and handle multivariate inputs, especially in complex or extended forecasting scenarios. Recent advancements in Transformer-based architectures, driven by self-attention mechanisms, offer a more effective approach for modeling global dependencies and complex temporal relationships. This paper compares traditional statistical methods and advanced neural models, including Sparse Attention, Transformer with Temporal Fusion, Artificial Neural Networks, and Block Hankel Networks. We evaluate these models under consistent experimental conditions using a real-world energy consumption dataset, applying performance metrics such as RMSE, MAE, and R-squared. Our results show that Transformer-based models and Sparse Attention architectures achieve nearly equivalent performance, with Transformers having a slight edge in scalability and long-sequence forecasting accuracy. These findings position Transformer-based models as a robust and scalable solution to the current challenges in energy-related time series prediction.