Enhancing Medium Term Wind Power Forecasting Accuracy With Dual Stage Attention Based TCN-GRU Model and White Shark Optimization

This research introduces a creative strategy for addressing the challenges of wind power predicting, crucial for effective renewable energy integration into the power grid. We propose a dual-stage attention-based Temporal Convolutional Network and Gated Recurrent Units (ATCN-AGRU) method for predicting wind energy over the medium term. In the first stage, a local attention mechanism captures fine-grained details and local dependencies, while the second stage employs a global attention mechanism to emphasize broader context and long-range dependencies within the data sequence. Our experimentation employs wind turbine data at 1-hr resolution, examining various time horizons from 24 hr to 1 week to assess multi-step forecasting precision and computational efficiency. Through rigorous statistical assessments, we demonstrate the model's validity, with the mean absolute percentage errors (MAPE) consistently below 9% for week-ahead forecasting. Model parameters were fine-tuned using white shark optimization (WSO), enhancing convergence and overall performance. The proposed hybrid model significantly outperforms standard forecasting methods, achieving a maximum MAPE of 8.07% for week-ahead forecasting and a minimal 4.44% for day-ahead forecasting. To test the model's stability, we extend the experiment to month-ahead forecasting, with the mean absolute error (MAE) ranging from 2.42 to 6.2% for weekly predictions.