<p>Forecasting in power-grid cyber-physical systems is hindered by frequency-band aliasing and multi-scale coupling in nonstationary series. We propose VLTGF, a dual-branch heterogeneous framework that performs adaptive frequency decoupling and coordinated feature fusion. The trend branch integrates Variational Mode Decomposition (VMD) with a Long Short-Term Memory network (LSTM) to capture slow dynamics, while the fluctuation branch combines a dilated Temporal Convolutional Network (TCN) with a Bidirectional Gated Recurrent Unit (BiGRU) to model long-range dependencies and abrupt changes. A lightweight fusion module aggregates the heterogeneous representations, and a differentiated optimization layer adaptively sets module-level hyperparameters, reducing manual tuning and stabilizing band segmentation and feature extraction. Evaluations on four grid-relevant datasets show that VLTGF attains state-of-the-art accuracy on three datasets and ranks second on the remaining one, outperforming Transformer variants (Informer, Autoformer, PatchTST) and recent hybrid baselines (VMD–CNN–GRU, VMD–TCN–GRU). Representative gains include a reduction over PatchTST on Temperature of 17.6% (RMSE), 15.1% (MAE), and 16.8% (MAPE), and an improvement over Informer on Power-consumption of 9.4%, 16.1%, and 8.9% on the same metrics. Holm-adjusted pairwise tests confirm the superiority on three datasets (p_adj &lt; 0.001). These results indicate that adaptive VMD configuration, dual-branch modeling, and the optimization layer jointly mitigate frequency aliasing and multi-scale coupling, enabling robust, high-precision forecasting for power-grid applications.</p>