Hybrid temporal fusion transformer and quantum long short-term memory network for typhoon forecasting at a wind power farm

Typhoons pose a significant risk to wind power generation and power grid stability, exacerbating wind resource losses and causing substantial economic damage. Therefore, accurate typhoon prediction is crucial for both onshore and offshore wind farms. In this study, a novel forecasting framework, termed the Temporal Fusion Transformer–Quantum Long Short-Term Memory (TFT-QLSTM) network, is proposed to enhance forecasting performance for wind farms, particularly those located in Gaomei, Taichung, Taiwan. The QLSTM component is based on variational quantum circuits (VQCs), which replace classical linear transformations. The proposed framework generates joint 24-h-ahead forecasts across seven typhoon and wind farm variables, optimized via Chaos Quantum Energy Valley Optimization (CQEVO) and interpretable through the TFT's variable selection network. Quantitatively, for the highly volatile wind gust at Gaomei, the proposed method reduces the MAE by approximately 31% compared with the hybrid baseline (TFT-LSTM) and by more than 77% compared with the standard Extreme Learning Machine. These improvements are averaged over 20 typhoon test events and are statistically significant according to the Diebold–Mariano (DM) test. Furthermore, the ablation study demonstrates that the performance gains are attributable to the QLSTM architecture rather than merely an increase in the number of model parameters. NVIDIA CUDA-Q is employed to enable GPU-accelerated simulation of the quantum circuits within the QLSTM component, with quantum circuit execution performed on an NVIDIA A100 Tensor Core GPU.