Wind Prediction under Random Data Corruption (Student Abstract)

Conner Flansburg, Dimitrios I. DiochnosAAAI 2022 · DOI

Abstract

In real-world meteorological applications, training data is frequently corrupted by hardware failures or transmission errors. This paper investigates the resilience of wind prediction models—specifically Lasso, Ridge, and Neural Networks—when corrupted data is identified and discarded. Our findings indicate that Lasso regression is exceptionally robust, maintaining high accuracy even with significant data loss, and often outperforms more complex neural network architectures in these scenarios.

Introduction

Data integrity is a persistent challenge in environmental sensing. Sensors fail, storage systems crash, and data can be intentionally manipulated. While significant research has addressed robustness in image classification, regression problems like wind speed prediction remain under-explored. This study addresses a practical scenario: if a system can detect corrupted data (e.g., via checksums), is discarding that data a viable strategy? We evaluate how much random data corruption can be sustained before predictive models for wind intensity become unreliable.

Methodology

We compared three modeling approaches:

  1. Ridge Regression: Linear regression with $L_2$ regularization.
  2. Lasso Regression: Linear regression with $L_1$ regularization, known for feature selection and sparsity.
  3. Multi-layer Neural Network: A non-linear model representing higher complexity.

Data corruption was simulated by randomly permuting training data and training models on progressively smaller subsets. This mimics the effect of discarding "useless" corrupted examples. Performance was measured using Root Mean Squared Error (RMSE), smoothed over 10 iterations to ensure consistency.

Experimental Setup

The study utilized a meteorological dataset of 16 cities (McGovern et al., 2021). Each city dataset contains approximately 3,140 to 3,200 instances with 2,723 to 3,567 features. Models predicted wind speed on a continuous scale or the Beaufort scale. An 80/20 chronological split was employed to simulate real-world forecasting conditions.

Results and Analysis

Lasso regression emerged as the most robust model. It achieved the lowest error rates when full data was available and demonstrated superior stability as data was removed.

Figure 1: Error Score vs. Training-Set Size for Ridge, Lasso, and Neural Network models. Lasso demonstrates superior stability and lower error as data is removed.

As shown in Figure 1, the Lasso error curve flattens much earlier than those of Ridge or the Neural Network. This indicates that Lasso reaches its performance plateau with significantly less data. For instance, in the city of Cheyenne (kcys), discarding one-third of the dataset resulted in only a 6% increase in RMSE. The inherent $L_1$ penalty in Lasso likely contributes to this robustness, providing a level of protection against data perturbations that the $L_2$ penalty or complex neural architectures do not match in this context.

Conclusion

Discarding corrupted data is a viable strategy for maintaining model performance in wind prediction. Lasso regression, in particular, offers an optimal balance of accuracy, stability, and interpretability. Future work will investigate targeted poisoning attacks and "certified robustness" guarantees to further secure these critical meteorological models against malicious interference.

Summary updated

BibTeX
@inproceedings{flansburg2022wind,
  author = {Conner Flansburg and Dimitrios I. Diochnos},
  title = {{Wind Prediction under Random Data Corruption (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21609},
}
Figure