Time-series forecasting, as a core technology supporting intelligent decision making, has important application value in fields such as power dispatch, traffic control, and economic early warning. However, existing methods still face the challenges of insufficient prediction accuracy when dealing with nonlinearity, multi-scale characteristics, and multivariate coupling relationships of complex time series. This paper focuses on key issues in multivariate time series forecasting and proposes an innovative algorithmic framework. To address the scale heterogeneity problem in multivariate time series, a dual-stream network architecture is proposed: first, since the seasonal and trend components exhibit inconsistent temporal characteristics, Long Short-Term Memory (LSTM) networks and linear layers are employed to capture the nonlinear fluctuations of seasonal components and the linear evolution patterns of trend components, respectively; second, a cross-scale attention mechanism is innovatively introduced to achieve adaptive fusion of multi-scale information by quantifying feature correlations across different time granularities. Finally, experiments on seven public datasets and the Henan Province load dataset demonstrate that the proposed algorithm outperforms existing mainstream methods in both prediction accuracy and robustness.