Hybrid machine learning approach for improved short-term PV power forecasting accuracy

• Sustainable Development Focus: The work emphasizes the necessity for sustainable development and the adoption of renewable energy sources, particularly photovoltaic (PV) solar power, in response to the growing global energy demand. • Short-Term Forecasting Methodology: A novel forecasting methodology is developed for short-term PV power generation, utilizing a hybrid model that combines Temporal Convolutional Networks (TCNs) with bidirectional layers of Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU). • Improved Accuracy and Efficiency: The proposed model significantly enhances forecast accuracy and operational efficiency, which are critical for real-time energy management and planning. • Data-Driven Seasonal Analysis: The research addresses traditional seasonal forecasting limitations by implementing a data-driven approach that leverages actual meteorological data to identify seasonal transitions, improving the understanding of seasonal dynamics. • Adaptive Seasonal Segmentation: By analyzing change points in key climate variables, the methodology moves away from fixed date methods, allowing for a flexible model that adapts to climate variability. • Strategic Transfer Learning: The study highlights the importance of strategic transfer learning in improving short-term PV power predictions, achieving substantial accuracy gains even in adverse weather conditions. The necessity for sustainable development and the adoption of renewable energy sources, notably photovoltaic (PV) solar power, is underscored by the expanding global need for energy. The integration of solar electricity into current energy networks creates challenges despite its tremendous potential to reduce environmental issues. This is especially true given its intermittent nature, which is influenced by meteorological circumstances. Precise short-term forecasting is important to assure grid stability and proper energy management in light of this volatility. Our work provides a forecasting method utilizing a hybrid model architecture that integrates bidirectional layers of Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM), and Efficient Channel Attention (ECA) with Temporal Convolutional Networks (TCNs). Forecast accuracy is significantly increased by this model, which successfully takes into account both long-term dependencies. With a remarkable configuration of TCN ECANet STACKED BiLSTMBiGRU, our model performs remarkably well when evaluated on the DKASC dataset from Uluru, Australia. It achieves an average RMSE of 0.028975, MAE of 0.0119275, and R² Score of 0.98832. This study extends the capabilities for precise solar power forecasting and helps to strategically incorporate renewable energy sources into the grid, along with more general sustainability goals.