FP-TTFNet: fourier periodic embedding enhanced multi-scale deep learning for monthly runoff prediction at contrasting hydrological stations
Reliable monthly runoff forecasts are essential for reservoir operation, water allocation, and flood–drought preparedness, but their robustness across contrasting basins is limited by multiscale periodicity, nonlinear catchment response, and spatiotemporal heterogeneity. This study develops FP-TTFNet, an end-to-end forecasting framework that couples explicit periodic priors with adaptive nonlinear mapping and collaborative local–global dependency modeling for monthly runoff prediction. Fourier periodic embedding encodes seasonal and interannual phase information; a parametric nonlinear feature processor adapts to heterogeneous catchment responses; multiscale temporal convolution and Transformer self-attention jointly capture short-term fluctuations and long-range dependencies; and an operationally weighted hybrid loss emphasizes peak flows and long-horizon stability. The framework is evaluated on three climatically distinct stations—monsoon-dominated Manwan, arid Jiayuguan, and alpine-transition Hongshanhe—against LSTM, GRU, RNN, TCN, and Transformer using ten repeated runs and significance testing. FP-TTFNet consistently achieves the lowest errors and highest efficiency: at Manwan, MAE is reduced by 22.0% to 176.06 m 3 /s with KGE = 0.9591; at Jiayuguan, RMSE is reduced by 43.1% to 2.99 m 3 /s; at Hongshanhe, NSE reaches 0.9508 despite the 117 km 2 catchment. Across all sites, NSE exceeds 0.92, and KGE exceeds 0.95. The results show that integrating interpretable periodic priors, nonlinear adaptation, multiscale dependency fusion, and operation-oriented optimization provides robust monthly runoff forecasts for engineering decision-making.
Authors
- Wenchuan Wang (ORCID: https://orcid.org/0000-0003-1367-5886)
- Jia Li
- Dong-mei Xu
Institutions
- North China University of Water Resources and Electric Power (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1038/s41598-026-81184-8
- Primary Topic
- Hydrological Forecasting Using AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00