<i>H</i><sub>∞</sub> Filtering for Bias Correction in Post-Processing of Numerical Weather Prediction

Jaechan Lim, Hyung‐Min ParkJournal of the Meteorological Society of Japan Ser II 2019 · DOI

Abstract

In this paper, we propose an H-infinity ($H_\infty$) filtering approach for the prediction of bias in post-processing of model outputs and past measurements. This method adopts a minimax strategy that is a solution for zero-sum games. The proposed $H_\infty$ filtering approach minimizes maximum possible errors whereas a recently proposed approach that adopts Kalman filtering (KF) minimizes the mean square errors. The proposed approach does not need the information of noise statistics unlike the method based on the KF, while the training process is required. We show that the proposed approach outperforms the method based on the KF in experiments by applying real weather data in Korea.

Introduction

Numerical Weather Prediction (NWP) models are the backbone of modern forecasting, but they often exhibit systematic biases due to the complexity of atmospheric physics and computational constraints. These biases—differences between predicted values and actual ground observations—are particularly problematic in fields like renewable energy management. For instance, overestimating solar irradiance can lead power grid managers to under-prepare backup energy sources, risking instability. Traditionally, Kalman Filtering (KF) has been used to correct these biases. However, KF assumes that noise follows a specific Gaussian distribution and requires known statistical properties, which are rarely available for messy real-world weather data.

Methods

We propose using $H_\infty$ (H-infinity) filtering for bias correction. While a Kalman Filter attempts to minimize the Mean Square Error (average mistake), the $H_\infty$ filter adopts a minimax strategy derived from game theory. This approach treats bias correction as a zero-sum game where the filter tries to minimize the maximum possible error that "Nature" (noise) can cause.

Flowchart of the bias correction process using NWP and observations.

Technically, the bias is modeled as an $n$-th degree polynomial of the NWP forecast. The $H_\infty$ filter recursively updates the coefficients of this polynomial. The primary advantage is robustness: the filter does not need to know the statistical "shape" of the noise to work effectively; it only requires the noise to be bounded.

Experimental Setup

The study utilized real-world meteorological data from the Korean Meteorological Administration (KMA), focusing on downward short-wave radiation flux (horizontal irradiance). Data was collected from five cities in South Korea—Chuncheon, Seosan, Jeonju, Jinju, and Andong—covering periods between March and December 2014. The proposed $H_\infty$ method was benchmarked against raw NWP outputs and a state-of-the-art Kalman Filter implementation.

Key Results

Experimental results demonstrate that the $H_\infty$ filter consistently outperforms the KF in terms of both Mean Absolute Error (MAE) and Maximum Absolute Error. For 24-hour forecasts in Chuncheon, $H_\infty$ reduced the MAE by approximately 17% compared to raw NWP, while the KF achieved a 15.5% reduction.

Comparison of absolute error over 30 days for raw NWP, KF, and H-infinity filter (h=24).

A critical finding is the $H_\infty$ filter's ability to cap "big misses." In one scenario, the KF's worst error was 0.6218 mW/m², while the $H_\infty$ filter limited the maximum error to 0.4104 mW/m². Furthermore, the $H_\infty$ algorithm proved to be computationally simpler and faster to execute than the adaptive KF, making it highly suitable for real-time operational forecasting.

Conclusion

The $H_\infty$ filtering approach provides a robust and efficient alternative for NWP bias correction, particularly when noise statistics are unknown. While it requires a training process to tune internal parameters (such as the performance bound $\gamma$), its superior performance in minimizing maximum errors makes it a valuable tool for critical forecasting applications. Future research will explore applying this method to other variables like wind speed and developing automated parameter tuning techniques.

Summary updated

Abstract

In this paper, we propose an H-infinity (H∞) filtering approach for the prediction of bias in post-processing of model outputs and past measurements. This method adopts a minimax strategy that is a solution for zero-sum games. The proposed H∞ filtering approach minimizes maximum possible errors whereas a recently proposed approach that adopts Kalman filtering (KF) minimizes the mean square errors. The proposed approach does not need the information of noise statistics unlike the method based on the KF, while the training process is required. We show that the proposed approach outperforms the method based on the KF in experiments by applying real weather data in Korea.
BibTeX
@article{lim2019lt,
  author = {Jaechan Lim and Hyung‐Min Park},
  title = {{\<i\>H\</i\>\<sub\>∞\</sub\> Filtering for Bias Correction in Post-Processing of Numerical Weather Prediction}},
  journal = {Journal of the Meteorological Society of Japan Ser II},
  year = {2019},
  doi = {10.2151/jmsj.2019-041},
}
Figure