iTBG-Net: Enhancing Short-Term Renewable Energy Forecasting with Incremental Deep Learning

Intelligent microgrids rely on state-of-the-art time series forecasting strategies, such as Machine Learning (ML), Deep Learning (DL), and attention-based models, to predict energy prices, demand responses, and short-, medium-, and long-term loads. These models take into account trends, seasonality, cyclic patterns, and noise, which face challenges when dealing with smart grid energy data due to factors such as real-time variability (customer consumption behaviour), high-frequency sampling (collected in seconds or minutes), and intermittency in renewables (dependence on weather patterns). To make a significant i mpact o n t his s ituation, i ncremental o nline learning models incorporate real-time adaptability from data fluctuations and dynamics while simultaneously retaining and adopting new knowledge. These models also handle high-frequency sampling by learning from small streams of data in real time rather than relying on batch-based learning. This article introduces the Incremental Temporal BiGRU Network (iTBG-Net), which combines a Temporal Convolutional Network (TCN), Bidirectional GRUs, and replay buffers to make it easier to retrieve previous information. Our proposed model outperforms the benchmark models: incremental CNN, LSTM, and XLSTM. We achieved a training loss of 0.0066 and a validation loss of 0.0190, as well as a training accuracy of 0.8128 and a validation accuracy of 0.8005. This makes it a powerful tool to improve forecasting accuracy in rapidly changing environments. Our open-source code is available on GitHub: (https://github.com/Aftab-Hussain302/iTBG-Net-oneline-Energy-forecasting.git).