آيات جهاد

Abstract High-consequence decision-making hinges on the reliable quantification an‎d transparent visualization of temporal uncertainty in dynamic systems. Conventional time-series forecasting models often produce prediction intervals that are poorly calibrated, leading to epistemic opacity an‎d compromising the integrity of risk communication required for robust governance an‎d "Stakeholder updat‎es." This thesis, Uncertainty Unveiled: Exploring Techniques for Visualizing Uncertain Data, addresses this deficiency by developing an‎d validating a novel CRPS-Tuned Hybrid Probabilistic Ensemble framework. The core innovation lies in the systematic fusion of complementary models—specifically Long Short-Term Memory (LSTM) networks for capturing complex temporal dependencies an‎d Mixture Density Networks (MDNs) for generating full predictive probability distributions. The architecture also incorporates adaptive non-parametric estimators an‎d a zero-inflation component to robustly model the stochastic variability inherent in real-world time-series. The framework was rigorously validated using highly volatile, domain-specific meteorological time-series data. Model fusion an‎d optimization were driven by the Continuous Ranked Probability Score (CRPS) to guarantee mathematically balanced sharpness (accuracy) an‎d reliability (calibration). The empirical validation demonstrated the modelʹs superior performance, achieving well-calibrated uncertainty with a Prediction Interval Coverage that accurately aligned with the nominal target (e.g., 80.83% coverage for an 80% interval). This research provides a scientifically validated methodology for transforming raw, uncertain data into trustworthy, decision-ready visualizations. By ensuring the statistical integrity of the uncertainty bounds, the thesis significantly advances the stan‎dards for epistemic transparency an‎d forecast reliability, offering a crucial contribution to the development of robust, uncertainty-aware decision support systems in critical temporal domains. So, by unveiling uncertainty rather than concealing it, the study enables a shift toward more transparent, reliable, an‎d actionable data-driven decision-making.

(تجسم عدم قطعيت، مدل‌هاي احتمالي تركيبي، كاليبراسيون پيش‌بيني، پيش‌بيني سري‌هاي زماني، پوشش بازه پيش‌بيني، امتياز احتمال رتبه‌بندي‌شده پيوسته (CRPS)، شبكه‌هاي چگالي تركيبي (MDN)، مديريت ريسك و LSTM (حافظه كوتاه‌مدت بلندمدت).)

(Uncertainty Visualization, Hybrid Probabilistic Models, Forecast Calibration, Time-Series Forecasting, Prediction Interval Coverage, Continuous Ranked Probability Score (CRPS), Mixture Density Networks (MDN), Risk Management, an‎d LSTM (Long Short-Term Memory).)

ayat Jihad

Dr. Mohammad Reza Kangavari