ybrid singular spectrum analysis (SSA) with channel-independent patch time series transformer (PatchTST) optimized by multi-agent HPO for time series forecasting
Abstract
Accurate forecasting of thermal power generation is essential for ensuring grid stability, reducing operational costs, and supporting sustainable energy management. This paper proposes a hybrid forecasting model that integrates Singular Spectrum Analysis (SSA) with a channel-independent Patch Time Series Transformer (PatchTST) to improve predictive performance in complex and noisy thermal power data. The SSA module filters the input signals, enhancing trend representation and reducing noise, while the transformer architecture effectively captures nonlinear temporal dependencies and long-range interactions. The proposed model (SSA-PatchTST), is hypertuned by a multi-agent hyperparameter optimization (HPO) strategy to ensure its best use. The proposed approach was evaluated on real thermal generation data over multiple forecasting horizons (5, 10, and 25-step ahead) and benchmarked against traditional statistical models and state-of-the-art deep learning methods. Results demonstrate performance gains, with reductions of more than 50\% in RMSE, MAE, and MAPE compared to the best baseline models. These findings highlight the potential of the proposed optimized SSA-PatchTST model as a robust and scalable solution for power forecasting, offering significant contributions to decision support in modern energy systems.