Publication

Abstract : This paper proposes a novel imbalance volume forecasting and subsequent balancing market operation framework that integrates decomposition, denoising, deep learning, and dynamic intraperiod ramping to improve forecasting accuracy and achieve cost-effective and reliable balancing market operation. A Robust Spectral-aware Trend–Seasonal Filtering (RSTSF) method is developed to jointly estimate trend, seasonal, and residual components under abrupt trend shifts and irregular seasonal patterns using robust likelihood modeling, structural coupling, and entropy-guided seasonal selection. To mitigate noise within the residuals, an Empirical Wavelet Transform combined with Piecewise Linear Approximation (EWT-PLA) is used for denoising. It effectively removes irrelevant noise while preserving the structural integrity of residual components. The Bidirectional Gated Recurrent Units (BiGRU) are used to forecast the trend and denoised-residual components due to their capability for bidirectional sequence learning. The Seasonal Attention BiGRU (SA-BiGRU) is proposed for seasonal components forecasting to capture periodic patterns. The parameters/hyperparameters of the proposed forecasting framework are optimized using a newly developed Newton–Raphson-Based Optimizer with Space Transformation Search (NRBO-STS), which reduces the risk of premature convergence and balances the exploration and exploitation capability. Furthermore, dynamic intraperiod ramping is incorporated into balancing market operation to ensure that Balancing Reserves (BRs) are practically deliverable in real-time, thereby improving market reliability and preventing BR shortfalls. A seasonal analysis conducted on the Belgian and Austrian power markets shows that the proposed framework achieves the lowest balancing costs, yielding savings of 23.31% and 25.58% compared to the second-best model in Belgium and Austria, respectively.