[Paper Review] Space-time calibration of wind speed forecasts from regional climate models

Luiz E. S. Gomes, Thaís C. O. Fonseca+2 more|arXiv (Cornell University)|Sep 27, 2019

demographic modeling and climate adaptation4 citations

TL;DR

This paper proposes a space-time calibration framework for wind speed forecasts from regional climate models using trans-Gaussian random fields and data augmentation to address temporal bias and skewed, censored wind speed distributions. The method improves forecast reliability and accuracy—particularly in complex terrain—by integrating dynamic, spatially structured Bayesian models with probabilistic scoring and model selection criteria.

ABSTRACT

Numerical weather predictions (NWP) are systematically subject to errors due to the deterministic solutions used by numerical models to simulate the atmosphere. Statistical postprocessing techniques are widely used nowadays for NWP calibration. However, time-varying bias is usually not accommodated by such models. Its calibration performance is also sensitive to the temporal window used for training. This paper proposes space-time models that extend the main statistical postprocessing approaches to calibrate NWP model outputs. Trans-Gaussian random fields are considered to account for meteorological variables with asymmetric behavior. Data augmentation is used to account for censuring in the response variable. The benefits of the proposed extensions are illustrated through the calibration of hourly 10 m wind speed forecasts in Southeastern Brazil coming from the Eta model.

Motivation & Objective

• Address systematic biases in regional climate model wind speed forecasts due to terrain and surface heterogeneity.
• Account for time-varying bias and seasonality in wind speed predictions through dynamic modeling.
• Handle the asymmetric and censored nature of wind speed data using trans-Gaussian random fields and data augmentation.
• Improve probabilistic forecast reliability by incorporating spatial dependence and temporal dynamics in statistical postprocessing.
• Evaluate model performance using proper scoring rules and Bayesian model selection criteria (DIC, LPML) for robust calibration.

Proposed method

• Employ spatiotemporal dynamic linear models (ST-DLMs) to model wind speed forecasts with time-varying mean and variance.
• Use trans-Gaussian random fields to model non-Gaussian, skewed wind speed distributions by applying a Box-Cox transformation.
• Apply data augmentation to handle censored observations (e.g., wind speeds below detection threshold) in the likelihood function.
• Implement Bayesian inference via Markov Chain Monte Carlo (MCMC) sampling to estimate model parameters and predictive distributions.
• Integrate spatial correlation structures using covariance functions that account for geographic distance and orographic complexity.
• Use probabilistic scoring rules (IS, CRPS) and model selection criteria (DIC, LPML) to compare and select optimal calibration models.

Experimental results

Research questions

• RQ1 How does incorporating time-varying bias and spatial dependence improve the reliability of regional climate model wind speed forecasts?
• RQ2 To what extent does data augmentation for censored wind speed observations enhance calibration performance compared to standard approaches?
• RQ3 Can trans-Gaussian random fields effectively model the skewed distribution of 10-meter wind speed in complex terrain?
• RQ4 How do different training window lengths affect calibration performance, and can the proposed model mitigate this sensitivity?
• RQ5 Which model selection criteria (DIC, LPML) best identify the most accurate and reliable calibration model for wind speed forecasts?

Key findings

• The proposed space-time calibration model significantly reduces mean absolute error (MAE) and root mean squared error (RMSE) in 24-hour wind speed forecasts compared to raw model outputs.
• The model with data augmentation and trans-Gaussian transformation outperforms standard approaches in probabilistic forecast accuracy, as measured by the interval score (IS) and continuous ranked probability score (CRPS).
• Bayesian model selection using DIC and LPML consistently favored the full space-time model with data augmentation, indicating superior fit and predictive performance.
• The model effectively captures seasonal and spatial variations in wind speed, particularly in regions with complex orography, reducing systematic under- or over-prediction.
• Residuals from the calibrated model show improved normality in Q-Q plots, indicating better calibration and reduced bias across seasons.
• The use of dynamic, time-varying parameters allows the model to adapt to changing meteorological conditions, reducing the sensitivity to fixed training window lengths.