Abstract: Response prediction is an effective means to accurately assess the operational condition of bridges. Due to variable service environments and complex mechanical mechanisms, bridge response monitoring data may have multiple characteristics, such as nonlinearity or non-Gaussianity. Traditional data-driven prediction models typically rely on feature simplification or data smoothing to extract the primary components of these complex data. This treatment often results in inadequate capture of the spatial patterns and local time-varying characteristics in the data. To address this issue, this paper proposes a prediction method for bridge structural responses using Swarm Decomposition (SWD) technology and the Patch Time Series Transformer (PatchTST) model. By effectively capturing key spatiotemporal features in the response data using SWD, the limitations of traditional time-frequency analysis techniques in analyzing complex internal signal structures are overcome. Furthermore, by integrating a multi-scale patch nesting mechanism, the ability of the model to represent local time-varying features is enhanced, thereby achieving accurate dynamic response prediction and improved generalization performance in complex operational environments. The effectiveness of the proposed method is validated through four measured data from two bridges. The results demonstrate that, compared with PatchTST, the proposed method reduces the value of Mean Squared Error (MSE) by 50% to 90%; compared with other Transformer architecture models (such as Autoformer and FEDformer), the MSE value is reduced by 10% to 60%. Compared with prediction models based on feature simplification techniques, the proposed method improves accuracy by over 30%.