A machine learning framework for radon forecasting with uncertainty quantification in the Campi Flegrei volcanic area

This study presents a machine learning framework for short-term forecasting of radon activity concentration in the Campi Flegrei volcanic area. We integrate fourteen years of radon, meteorological, and seismic observations through a dedicated preprocessing pipeline that addresses missing values, uneven sampling, and aggregation into 3-hour intervals. Two model families are compared: a tree-based predictor (XGBoost) and a sequence-based predictor (LSTM). Conformal prediction is employed as an uncertainty layer to complement deterministic forecasts with predictive intervals. Point-prediction performance is quantified in terms of Root Mean Square Error (RMSE) and Pearson correlation: the best one-step predictor achieves a normalised RMSE of about 15% (normalised by the mean concentration) together with a high correlation coefficient (ρ≈0.95). Uncertainty quantification relies on Conformal Prediction (CP), and is summarised by the empirical Prediction-Interval Coverage Percentage (PICP) with respect to the nominal level, together with interval sharpness measures; for a representative configuration, PICP remains close to the nominal level (around 90%). Finally, recursive free-run experiments highlight the expected horizon-dependent degradation: errors grow with the forecast horizon for both predictors, with XGBoost attaining slightly lower mean error while LSTM shows greater robustness through reduced variability across rollouts at medium and long horizons. Overall, the proposed framework integrates well-established methodologies for radon forecasting and uncertainty-aware evaluation, enhancing the reliability of environmental time-series forecasting and providing a reproducible tool.