MMWSTM-ADRAN+: Site-specific one-day-ahead maximum temperature forecasting in baghdad with classical, neural, and extreme-tail benchmarks

Accurate one-day-ahead prediction of daily maximum temperature is important for local heat-risk management, energy planning, and public-health preparedness. This revised manuscript explicitly frames MMWSTM-ADRAN+ as a site-specific statistical post-processing and autoregressive forecasting framework for a blended numerical weather prediction (NWP)-derived daily product for Baghdad, Iraq (2019–2024), rather than as a general climate-scale forecasting system. At prediction time, the model uses a 30-day historical window of variables available up to day t, including lagged daily maximum temperature and causal derived features; it does not use future observations or raw NWP ensemble forecasts for day t + 1. In response to the reviewers, the evaluation now includes persistence, seasonal climatology, AR(1)+seasonal linear regression, and multiple linear regression baselines, together with the neural baselines Temporal Transformer, TCN, and N-BEATS. On the held-out test window (2023–10–26–2024–12–31), MMWSTM-ADRAN+ achieved RMSE = 1.706 °C, MAE = 1.312 °C, R² = 0.973 and RMSE skill = 23.8% relative to persistence. The multiple linear regression baseline achieved the lowest overall RMSE (1.384 °C), showing that much of the predictability arises from lagged thermal persistence and engineered predictors. MMWSTM-ADRAN+ nevertheless achieved the lowest hot-tail RMSE among the reported models (1.172 °C). The manuscript therefore presents the model as a competitive extreme-tail diagnostic architecture, not as a universally superior climate-forecasting system. Claims are restricted to a single site, one variable, one-day lead time, and a blended NWP-derived reference product; independent station-observation validation remains a necessary next step.