[Paper Review] Forecast-Aware Model Driven LSTM

Sophia Hamer, Jennifer Sleeman+1 more|arXiv (Cornell University)|Mar 23, 2023

Air Quality Monitoring and ForecastingEnvironmental Science3 citations

TL;DR

This paper proposes a forecast-aware, cluster-based bidirectional LSTM model for bias correction in operational air quality forecasting, using unsupervised clustering of station geography and urbanization to partition data and train region-specific LSTMs. The method reduces ozone forecast RMSE by 27% (3.14 ppb) compared to the base model, with further improvements when including elevation in clustering.

ABSTRACT

Poor air quality can have a significant impact on human health. The National Oceanic and Atmospheric Administration (NOAA) air quality forecasting guidance is challenged by the increasing presence of extreme air quality events due to extreme weather events such as wild fires and heatwaves. These extreme air quality events further affect human health. Traditional methods used to correct model bias make assumptions about linearity and the underlying distribution. Extreme air quality events tend to occur without a strong signal leading up to the event and this behavior tends to cause existing methods to either under or over compensate for the bias. Deep learning holds promise for air quality forecasting in the presence of extreme air quality events due to its ability to generalize and learn nonlinear problems. However, in the presence of these anomalous air quality events, standard deep network approaches that use a single network for generalizing to future forecasts, may not always provide the best performance even with a full feature-set including geography and meteorology. In this work we describe a method that combines unsupervised learning and a forecast-aware bi-directional LSTM network to perform bias correction for operational air quality forecasting using AirNow station data for ozone and PM2.5 in the continental US. Using an unsupervised clustering method trained on station geographical features such as latitude and longitude, urbanization, and elevation, the learned clusters direct training by partitioning the training data for the LSTM networks. LSTMs are forecast-aware and implemented using a unique way to perform learning forward and backwards in time across forecasting days. When comparing the RMSE of the forecast model to the RMSE of the bias corrected model, the bias corrected model shows significant improvement (27\% lower RMSE for ozone) over the base forecast.

Motivation & Objective

• Address the challenge of increasing extreme air quality events, such as wildfires and heatwaves, which disrupt traditional linear bias correction methods.
• Improve operational air quality forecasting accuracy by correcting systematic biases in NOAA’s model outputs for ozone and PM2.5.
• Overcome limitations of monolithic deep learning models by partitioning training data using unsupervised clustering of geographic and urbanization features.
• Enable better generalization and sensitivity to regional and extreme events through region-specific, forecast-aware bi-directional LSTMs.
• Develop a scalable, modular deep learning framework compatible with operational forecasting systems and future high-resolution AI-driven models.

Proposed method

• Apply unsupervised K-means clustering to station features including latitude, longitude, elevation, and urbanization to group geographically and climatologically similar monitoring sites.
• Train a separate bi-directional LSTM for each cluster, enabling localized learning while maintaining global consistency in the forecasting task.
• Implement a forecast-aware training strategy that processes sequences both forward and backward in time, capturing long-term temporal dependencies.
• Use a triad architecture where the LSTM inputs are forecast intervals and outputs are corresponding ground-truth observations from AirNow stations.
• Optimize cluster count (K) via empirical evaluation, with K=25 yielding the best performance across ozone and PM2.5 forecasts.
• Compare LSTM performance with GRU-based variants to assess trade-offs between training time and prediction accuracy.

Experimental results

Research questions

• RQ1 Can a cluster-based, forecast-aware bi-directional LSTM improve bias correction in operational air quality forecasting compared to traditional statistical methods?
• RQ2 How does incorporating geographic and urbanization features into clustering affect the model’s ability to generalize and reduce forecast error?
• RQ3 What is the optimal number of clusters (K) for balancing model complexity and performance in regional air quality forecasting?
• RQ4 Does the forecast-aware training mechanism enhance performance on extreme air quality events where signals are weak or absent?
• RQ5 How does the LSTM-based approach compare in performance and training efficiency to a single-station model and GRU-based alternatives?

Key findings

• The cluster-based forecast-aware LSTM reduced ozone forecast RMSE by 3.14 ppb (27% improvement) compared to the base model, with K=25 yielding the best performance.
• Including elevation in clustering features produced the largest RMSE reduction (3.27 ppb) for ozone, outperforming combinations without elevation.
• The single-station LSTM model achieved the highest reduction (3.60 ppb) but required training 271 separate models, highlighting a trade-off between performance and scalability.
• Replacing LSTMs with GRUs reduced training time by 10 hours but decreased RMSE improvement by 0.07 ppb, indicating a performance-efficiency trade-off.
• PM2.5 forecasts showed a 1.945 ppb reduction in RMSE, demonstrating the method’s effectiveness across multiple pollutants.
• The model maintained strong performance on individual stations (e.g., Station 60793001) even when not explicitly trained on them, indicating effective generalization across clusters.