Abstract <p>Transformer-based models have shown strong performance in time-series forecasting, but they often struggle to capture temporal patterns across multiple scales and lack interpretability. We present a visual analytics framework for multi-scale patch decomposition in Transformer-based time-series forecasting, which integrates a novel forecasting model with an interactive visual analytics system. The model, called MSPT model, introduces a multi-scale temporal patching mechanism that decomposes time series into scale-specific components, and incorporates intra-patch, inter-patch, and inter-channel attention to capture rich temporal dependencies and improve forecasting accuracy. To further enhance interpretability, we develop MSPT-vis, a visual analytics system that reveals attention distributions, patching effects, and prediction errors through coordinated views, enabling human-in-the-loop exploration and model refinement. Extensive experiments on six public datasets demonstrate that the MSPT model outperforms three state-of-the-art baselines, while ablation studies validate the contributions of its core components. Expert evaluation further highlights the utility of MSPT-vis for diagnosing model behavior and guiding parameter configuration. The whole framework is available at: <a href="https://gitee.com/btbuvislab/">https://gitee.com/btbuvislab/</a> mspt.</p> Graphical abstract <p></p>