Climate-driven forecasting and reinforcement learning for Aedes albopictus control in Guangzhou, China: leakage-controlled benchmark, model results and trained PPO policies
Description
This dataset supports two companion studies on Aedes albopictus surveillance and control in Guangzhou, China. Part 1 (paper_a_forecasting): a leakage-controlled benchmark for vector-density forecasting. It includes ERA5 climate reanalysis for 10 districts (2015-2022), MOH-sourced Breteau index records, the mechanistic population-flow-index (PFI) simulator used as ground truth, and unified model results (SARIMA, Lasso, Random Forest, XGBoost, LSTM, Graph Attention Network) covering the urbanization-proxy leakage experiment, leave-one-grid-out validation, graph-structure sensitivity tests, SHAP feature importance, and double machine learning lag/dose-response analyses. Part 2 (paper_b_rl): a Markov decision process formulation of multi-zone proactive vector control with source reduction and space spraying. It includes the environment code, training/evaluation scripts, 18 trained PPO policy checkpoints (seeds and variants), and evaluation results against static threshold and calendar-based rules, including grid-tuned baselines, cost-weight sensitivity (50-fold), and parameter perturbation (plus/minus 30%). All results are provided as CSV files so every figure and number in both manuscripts can be reproduced without retraining. Documentation is in the README of each subfolder. Related repositories: code and live updates at <https://github.com/GustaveKateb0709/aedes-benchmark-guangzhou> ; long-term archive DOI 10.5281/zenodo.21948783.
Files
Institutions
- Guangdong Peizheng CollegeGuangdong, Guangzhou