Source-linked AI summary
Enhanced Artificial Neural Networks Using QHAdamW in Air Quality Forecasting
Mary Joy Daniel Vinas
TL;DR
Air quality forecasting in the Philippines has received limited attention, particularly for convergence and optimizer innovation in ANN models. This study proposes QHAdamW, combining QHAdam and AdamW, and reports improved forecasting performance and convergence for PM2.5 and PM10 AQI prediction.
Problem
Limited Philippine air quality forecasting research and convergence constraints motivate evaluating newer optimizer innovations in ANN-based AQI prediction.
Method
A feed-forward ANN uses the hybrid QHAdamW optimizer, combining quasi-hyperbolic momentum with decoupled weight decay, to forecast PM2.5 and PM10 AQI separately.
Results
QHAdamW improved forecasting metrics and convergence, with PM10 predicting accuracy reported as 88.03%.
Takeaways & Limitations
QHAdamW is reported as a suitable drop-in replacement for Adam in ANN-based AQI forecasting in the Philippines.
Takeaways & Limitations
Parameter tuning requires substantial trial and error and repeated runs, creating a major computational-cost disadvantage.
Abstract
from arXiv · showhide
The study employed an Artificial Neural Network in combination with the optimized Adaptive Moment Estimation (Adam) algorithm, currently the only AQI forecasting model available in the Philippines. The modified QHAdamW - Quasi-Hyperbolic Momentum (QHAdam) and Adam with decoupled weight decay (AdamW) were both extensions of the Adam optimizer, and both offer unique advantages for training ANN. The proposed QHAdamW optimizer addresses the issues on convergence, generalization, and forecasting performance of Adam. Hyperparameter tuning results revealed that 0.01 and 0.001 were the most effective optimal values for the generalization performance of QHAdamW. The comparative analysis results using seven evaluation metrics revealed that the error value range is lower, and the regression coefficient, having a value approximately equal to 1, improved the model accuracy performance. Likewise, the model converges to a satisfactory level of performance with the convergence performance results of lower loss values as obtained from training and validation losses. Based on data from a real-time air quality tracking station in Manila, a feed-forward neural network is used to predict the AQI of PM2.5 and PM10 separately. This model can be used to forecast Particulate Matter (PM), to help the Department of Environment and Natural Resources-Environmental Monitoring Bureau (DENR-EMB) implement a comprehensive air quality management.
1. Introduction
Air-quality forecasting matters for public-health protection, yet Philippine practice has focused mainly on observation and reporting. The study proposes QHAdamW, combining QHAdam and AdamW within an ANN to address convergence, generalization, and AQI-forecasting gaps for PM2.5 and PM10.
- Motivation: 19 µg/m3 is the reported average annual Filipino exposure to air pollutants, 1.9 times the WHO recommended threshold.
- Motivation: Accurate short-term PM2.5 and PM10 AQI forecasts support early warnings, health advisories, and operational planning by local agencies.
- Research gap: Philippine air-quality practices primarily emphasize observation and reporting, while forecasting systems remain limited in convergence and methodological diversity.
- Research gap: ANNs can describe complex nonlinear interactions and generalize to new data, motivating their use for AQI forecasting.
- Proposed approach: QHAdamW combines quasi-hyperbolic momentum with decoupled weight decay to address convergence stability, local minima, and weight-decay limitations in existing optimizers.
- Proposed approach: The study asks whether QHAdamW outperforms traditional Adam for convergent, generalized, and accurate PM2.5 and PM10 AQI forecasting in urban Philippine settings.
2. Materials and Methods
The study develops a feed-forward ANN for air-quality forecasting and compares Adam with the proposed QHAdamW optimizer. QHAdamW combines quasi-hyperbolic momentum with decoupled weight decay, with tuning, training, and evaluation used to assess forecasting performance and convergence.
- Data and procedures: 14,328 hourly PM2.5 and PM10 observations from January 2020 to December 2024 were collected at an EMB-NCR monitoring station in Manila.The station used a Teledyne T640 PM analyzer for real-time air-quality monitoring.
- ANN architecture: The forecasting model was a feed-forward multilayer ANN implemented in PyTorch, using pollutant concentrations as input features.The architecture included an input layer, hidden layers, and an output layer.
- Optimizer design: QHAdamW combined QHAdam's quasi-hyperbolic momentum with AdamW's decoupled weight decay as a custom PyTorch optimizer.The method integrates momentum modification with separation of weight decay from gradient-based updates.
- Hyperparameter tuning: Learning rates, weight decay values, and epsilon constants of 0.1, 0.01, and 0.001 were evaluated through grid search, with each configuration run ten times.Results were averaged across different random seeds to assess stability and reproducibility.
- Evaluation: Forecasting performance and convergence were assessed with regression metrics, using lower error values to identify better-fitted models.The study compared the ANN with QHAdamW against the original Adam optimizer.
3. Results and Discussion
Across PM2.5 and PM10 forecasting, QHAdamW generally outperformed Adam in error, stability, validation performance, and AQI classification accuracy. Results were based on repeated runs and comparisons across multiple evaluation metrics.
- Hyperparameter tuning: 0.01 learning rate, weight decay, and epsilon produced the fewest PM2.5 errors, while PM10 performed best with a 0.1 learning rate and 0.001 weight decay and epsilon.
- Accuracy performance: QHAdamW consistently produced better results than Adam across all examined error metrics for both PM2.5 and PM10.The metrics included R2, RMSE, MSE, and MAE, among others.
- Stability: 10 independent runs showed consistent QHAdamW gains, with lower standard deviations and improvement rates ranging from 142.8% to 847.5%.The lower standard deviations indicated greater stability and consistency than Adam.
- Regression and error metrics: QHAdamW achieved evaluation values closer to a regression coefficient of 1 and the best accuracy and forecasting-performance values across RMSE, MAE, R2, MSE, NRMSE, MAPE, and WAPE.
- Convergence performance: QHAdamW reduced training loss by 45.88% and validation loss by 45.21% relative to Adam, indicating improved learning and generalization.For PM10, the reported validation-loss difference was 0.263, and QHAdamW provided a better learning/generalization balance.
- AQI classification accuracy: PM2.5 accuracy increased by 2.53%, from 51.71% with Adam to 53.05% with QHAdamW, while PM10 accuracy increased by 9.99%, from 79.24% to 88.03%.
4. Conclusions and Recommendations
QHAdamW improved ANN-based AQI forecasting for PM2.5 and PM10, with performance gains across multiple metrics and improved convergence. The study recommends broader pollutant applications and additional evaluation criteria.
- RMSE = 3.38%, MSE = 6.68%, MAE = 2.08%, and predicting accuracy = 53.05% improved for PM2.5 with QHAdamW.
- RMSE = 2.70%, MSE = 5.48%, MAE = 1.60%, and accuracy = 88.03% improved for PM10 with QHAdamW.
- Training losses decreased to 45.88% and 36.15% for PM2.5 and PM10, respectively, while convergence also improved.
- Future evaluation can expand QHAdamW to sulfur oxides and nitrogen oxides and include model speed among the assessment criteria.