In India, diversified energy sources are integrated with rapid growth in electricity consumption, hence demanding effective forecasting methods in the smart grid framework. With the increasing scale and complexity of energy distribution networks, highly accurate short-term load forecasting (STLF) models are required to optimize power generation and consumption. Traditional STLF methods rely on relatively smaller datasets and need more scalability, limiting their performance when energy consumption data grows exponentially. This research focuses on the optimization of renewable energy generation and consumption in smart grids using deep learning tailored for energy consumption patterns across different distribution transformers. It uses time-series load data, starting from thorough data cleaning to improve dataset quality and capture seasonal variations critical for predictive accuracy. Three deep learning architectures are used: Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Temporal Convolutional Network (TCN). To avoid overfitting during the training process, two different data sets are made. 80% of the data is used for training, whereas 20% is employed for evaluation. GRU gave the best accuracy score, which was 92.5%, MAE was 3.5%, and RMSE came out to be 4.2%. Further validation of the predictability is obtained from confusion matrices that analyze the models. With minimal false positives and negatives, the effectiveness of the GRU model is also established. Energy forecasting, which is significant for the efficient management and resource allocation of smart grids, can be improved with the help of deep learning techniques. This research contributes to the research of energy forecasting and carries practical implications for utilities in optimizing operations or reducing wastage toward sustainable energy.