ML Precipitation Prediction Framework
Manuel Ricardo Pérez Reyes, Marco Javier Suárez Barón, Oscar García-Cabrejo, Octavio Castillo-Reyes
Open-source framework for monthly precipitation prediction in mountainous areas using hybrid deep learning. The framework provides reference implementations for eight model families and a uniform training, evaluation, and benchmarking pipeline: ConvLSTM family — baseline, bidirectional, residual, and attention variants Fourier Neural Operators — pure FNO and FNO-ConvLSTM hybrid decoder Graph Neural Networks with Temporal Attention (GNN-TAT) — GCN, GAT, and GraphSAGE encoders combined with multi-head temporal attention Late Fusion meta-learner — Ridge regression combining ConvLSTM and GNN-TAT predictions The framework was developed and validated on the Boyacá (Colombian Andes) CHIRPS-SRTM dataset (61×65 grid, 518 monthly steps, 1981–2024), with multi-seed validation across seeds {42, 123, 456} and lead times H=1..12. The Late Fusion ensemble attains R²=0.655±0.018 across seeds; a per-cell monthly climatology (R²=0.73) is reported as an explicit predictability ceiling. Verified reference environment. Python 3.12.5, PyTorch 2.6.0+cu124, PyTorch Geometric 2.7.0, NVIDIA CUDA 12.4. The requirements-lock.txt file pins every transitive dependency for byte-level reproducibility. v1.0.1 scope. Documentation and citation-metadata release synchronized with the archived DOI: full author list, published-companion references, and the benchmark artifact accompanying the manuscript under submission to Earth Science Informatics. Establishes the eight-family benchmark, the BASIC/KCE/PAFC feature taxonomy, the multi-seed multi-horizon evaluation protocol, and the Late Fusion ensemble interface.