Physics-Guided Action-Conditioned Latent World Models for Risk-Aware Thermal Planning in 5G Radio Access Infrastructure-Reproducibility Package

Reproducibility package for the article "Physics-Guided Action-Conditioned Latent World Models for Risk-Aware Thermal Planning in 5G Radio Access Infrastructure" by J. O. Aji, P. Xiao and M. Shojafar (submitted to IEEE Access, 2026). OverviewThe package reproduces every synthetic-benchmark result in the paper: forecasting accuracy and calibration (Tables III, A1–A2; Figs. 3–4), hidden-state reconstruction (Fig. 5), counterfactual planning, ablations and paired statistics (Tables IV–V; Figs. 6–7), and site heterogeneity (Table VI). PG-WM embeds a two-node resistance–capacitance (RC) thermal transition with history-conditioned bounded parameters in an action-conditioned latent world model, adds a bounded recurrent residual and a heteroscedastic Gaussian head, and drives a risk-aware planner evaluated against simulator ground truth. Contents1. simulator/ – two-node RC digital twin of 5G RAN site thermal dynamics (Eqs. 1–5) with ambient, load, solar and outage drivers, cooling-health degradation, sensor faults (stuck-at, drift, spikes) and an operator-like behaviour policy.2. models/ – PG-WM and six baselines: persistence, nominal Physics-RC, Ridge-ARX, XGBoost, LSTM and GRU-WM.3. planning/ – 20-plan action set, upper-confidence feasibility test (z = 1.645), operating-cost model, simulator replay, regret and violation metrics (Eqs. 11–13).4. configs/ – all training and evaluation settings (Table I), seeds 0, 1, 2.5. data/synthetic/ – 48 in-distribution sites (75 days) and 20 out-of-distribution sites (30 days, +2.0 °C ambient, 1.8× degradation) at 5-min resolution (1,209,600 records), per-site simulator parameters (Table II) and the site split (31 train / 6 validation / 11 test).6. checkpoints/ – trained LSTM, GRU-WM and PG-WM models for seeds 0–2.7. logs/ablations/ – training logs for the physics-off and action-off variants.8. results/ – site-level forecasting metrics, decision-averaged planning summaries and paired-test outputs.9. figures/ – scripts and source data for all figures. How to reproduce(1) Install the dependencies in requirements.txt. (2) Regenerate the synthetic data, or use data/synthetic directly. (3) Train each model for seeds 0, 1 and 2. (4) Run the forecasting evaluation on the test and OOD splits. (5) Run the planning evaluation on the fixed 400 decision states per split. (6) Regenerate the figures. Exact commands are in README.md. Integrity checksCorrect outputs satisfy: (a) thousands of evaluation windows per site (about 3,600 per ID test site and 1,440 per OOD site at a 30-min stride); (b) RMSE_all² equals the mean of the twelve squared horizon RMSEs in every row; (c) 400 decision states × 3 seeds = 1,200 paired decisions per split, shared by all models. Ground-truth columns prefixed GT_ are for evaluation only and must never be model inputs. Known limitationsFrozen prediction tensors from the original runs were not archived. The archived RMSE_h12 field duplicates RMSE_all; 60-min RMSE was regenerated for the three neural models. End-to-end equivalence between reloaded checkpoints and archived predictions has not yet been demonstrated. All dynamics data are synthetic. Field telemetryThe operator telemetry (250 sites, 12,000 records) used for the steady-state consistency check in Section VIII is proprietary to the network operator and is not included. Aggregated results are provided; access to the underlying records may be requested from the corresponding author, subject to the operator's approval. LicencesCode: MIT. Data, checkpoints, tables and figures: CC BY 4.0. ContactJacob Onu Aji, MTN Nigeria Communications Plc – [email protected] – ORCID 0009-0007-2209-2371