Source-linked AI summary
Air quality prediction using optimal neural networks with stochastic variables
Ana Russo, Frank Raischel, Pedro G. Lind
TL;DR
Air-quality forecasting needs useful predictions without relying on large sets of input variables. The paper derives stochastic variables from multivariate measurements and uses them to train neural networks. It reports substantially fewer inputs with predictive power maintained, with reductions typically greater than a factor of two and reaching a factor of ten at large time-lags.
Problem
ANN-based air-quality forecasts can require large training datasets, motivating a way to reduce input information without significantly changing predictive power.
Source: §1 · pp. 1–2Method
The authors derive eigenvariables from coupled stochastic measurements by transforming coordinates using diffusion-matrix eigensystems, then use the derived variables as ANN inputs.
Source: §1 · p. 2, §2.2 · p. 4Results
Input data were reduced typically by more than a factor of two and by a factor of ten for large time-lags, while predictive power was maintained.
Source: §4 · p. 9, §5 · p. 10Takeaways & Limitations
The reduced-input approach is presented as a step toward faster air-quality predictions and more efficient alert systems, with possible adaptation to weather and geophysical forecasting.
Source: §5 · p. 11The procedure assumes that each measurement variable represents a property evolving according to a stochastic equation.
Source: §2.2 · p. 4
Abstract
from arXiv · showhide
We apply recent methods in stochastic data analysis for discovering a set of few stochastic variables that represent the relevant information on a multivariate stochastic system, used as input for artificial neural networks models for air quality forecast. We show that using these derived variables as input variables for training the neural networks it is possible to significantly reduce the amount of input variables necessary for the neural network model, without considerably changing the predictive power of the model. The reduced set of variables including these derived variables is therefore proposed as optimal variable set for training neural networks models in forecasting geophysical and weather properties. Finally, we briefly discuss other possible applications of such optimized neural network models.
1. Introduction
Urban air pollution forecasting matters for health and air-quality alerts, but common approaches simplify atmospheric processes and neural networks often require large training datasets. This paper applies derived stochastic variables as ANN inputs to reduce input requirements while maintaining predictive power.
1. Introduction
- Air-pollution forecasts are a priority for anticipating low-air-quality episodes and protecting urban residents, particularly sensitive groups.The passage highlights children and people with previous heart or respiratory insufficiency.Source: §1 · p. 1
- Straightforward models are quick to implement but simplify atmospheric processes, while more detailed approaches require extensive accurate inputs and substantial computation.The cited limitations include describing pollutant transport and chemical behavior, especially for secondary pollutants.Source: §1 · p. 1
- ANNs have been used for air-quality forecasting, but their need for large training datasets has remained a disadvantage.Prior applications include hourly averages and daily maxima.Source: §1 · p. 1
- Stochastic data-analysis methods previously derived reduced-fluctuation variables from NO2 measurements and were developed to estimate drift and diffusion in stochastic processes.The framework has also been applied to turbulent flows, climate indices, wind-turbine performance, stock-market indices, and oil prices.Source: §1 · pp. 1–2
- Derived stochastic variables used as ANN inputs reduce the amount of input data needed for a given accuracy while maintaining predictive power.The authors argue these variables incorporate temporal correlations between independent, spatially separated monitoring stations.Source: §1 · p. 2
2. Methods
The methods combine a feed-forward ANN framework with stochastic-data analysis that derives a smaller set of variables from coupled measurements. Diffusion-matrix directions are used to transform the variables, while data conditions and validation procedures guide their use.
2.1. The Neural Network framework
- The ANN uses feed-forward layers, with weighted inputs and bias passed through activation functions; the study uses an input layer and one hidden layer.The authors report that more layers were redundant in their tests and that minimizing hidden neurons can help generalization and avoid overfitting.Source: §2.1 · pp. 2–3
- ANN performance is assessed using four rotating year-based calibration-validation splits and averaged Pearson correlation, RMSE, and skill against persistence.The Levenberg-Marquardt method is used for training; persistence and linear regression are baseline models.Source: §2.1 · p. 3
2.2. Deriving optimal stochastic variables
- The stochastic-variable procedure assumes each measurement property evolves according to a stochastic equation, and derives drift and diffusion from the joint measurements.The measurements are treated as a state vector with independent Gaussian stochastic forces; conditional moments are estimated from small-time behavior, using the lowest-τ value of M(τ)/τ when a linear fit is not possible.Source: §2.2 · p. 4
- For the NO2 application, periodicities are filtered by detrending, and the data’s Markovian properties are checked as a prerequisite for the Langevin-process assumption.The authors note that alternative procedures may be used when Markovian properties are not observed.Source: §2.2 · p. 4, Figure 3 · p. 5
- The diffusion matrix is diagonalized through a coordinate transformation whose eigenvalues indicate stochastic-force amplitudes and eigenvectors indicate their directions.This decouples stochastic contributions across the transformed variables, allowing attention to focus on coordinates associated with larger eigenvalues when they differ substantially.Source: §2.2 · p. 4
3. Data
The study uses hourly NO2 measurements from two Lisbon stations and derives stochastic variables from their coupled behavior for ANN inputs. Its optimal model typically reduces input variables by more than half while producing similar performance to the standard model.
3.1. Target data
- Hourly NO2 data from Chelas and Avenida da Liberdade cover 2002–2006, with about 30,000 measurements and roughly 1% discarded.The stations are 4.8 km apart.Source: §3.1 · p. 5
3.2. Input data for ANN training
- The first four years support model construction and internal evaluation, while 2006 is used for independent evaluation.The procedure trains on three years and predicts the held-out year within 2002–2005, then uses the resulting weights and biases for 2006.Source: §3.2 · p. 5
- The two NO2 series are detrended to remove periodic cycles before deriving stochastic variables from their coupled drift and diffusion structure.Detrending uses 52-week and then one-day segments; the resulting series show no periodic patterns and satisfy Markov properties and a delta-correlated Gaussian-like noise signature.Source: §3.2 · p. 6
- Input variables in the optimal model are typically reduced by more than 50% versus the forward-stepwise model, reaching 90% at the largest time lags, while performance remains similar.There are three exceptions to the typical reduction; Table 3 reports similar results to the standard model using fewer inputs.Source: Table 2 · p. 7, §3.2 · p. 8, Table 3 · p. 9
- Forward stepwise regression adds predictors that reduce error significantly and stops when no further variable passes a partial F-test.Four ANN input sets are considered, including the standard set, sets with transformed variables, and the reduced optimal set.Source: §3.2 · p. 7
- The transformed variable V+ corresponds to the largest eigenvalue direction of the diffusion matrix and captures more of the coupled concentrations’ correlations than V−.The transformed variables are added to the ANN inputs alongside pollutant, meteorological, and other available variables.Source: §3.2 · pp. 6, 8
4. Stochastic variables as optimal input for neural networks
The optimal model uses a reduced set of inputs selected by correlation with the target, while generally retaining predictive power comparable to the standard model. Across reported measures, differences are small, with some exceptions at the 24-hour lag.
4. Stochastic variables as optimal input for neural networks
- Variables with correlation to the target equal to or greater than that of V+ are retained for the optimal model.The preceding section reports that V+ is more correlated with the target than most meteorological and air-quality variables.Source: §4 · p. 9
- Pearson correlation is almost identical between models except at 24 hours, while mutual information shows no deviations apart from one exception in each evaluation period.The independent-sample Pearson results are relatively lower than calibration-validation results, except for AL at 24 hours.Source: §4 · p. 10
- The decrease in skill against persistence is ≲4% for short lags, except at 24 hours; independent-sample results are higher at every lag.The comparison is between MO and the standard model.Source: §4 · p. 10
- RMSE is almost unchanged at shorter lags, but increases by ≲3% for MO at 24 hours; independent-sample RMSE is usually lower for MO, except for AL at 3 hours.These comparisons are between MO and the standard model.Source: §4 · p. 10
- With less than half the input data, model MO has predictive power at least as good as the standard model, and is generally comparable across measures.The authors propose MO as an alternative for air-quality prediction.Source: §4 · pp. 9–10
5. Conclusions
Derived eigenvariables reduce the ANN input requirement by more than a factor of two typically and by a factor of ten at large time lags, while preserving predictive power. They incorporate temporal correlations across independent, spatially separated monitoring stations and are proposed for faster air-quality prediction and adaptation to weather or geophysical forecasting.
5. Conclusions
- Input data are reduced by more than a factor of two typically and by a factor of ten at large time lags, while predictive power is maintained.The reduced inputs include derived eigenvariables.Source: §5 · p. 10
- The derived variables incorporate temporal correlations between independent, spatially separated monitoring stations.Source: §5 · p. 11
- In air-quality forecasting, reduced inputs enable faster predictions without significantly changing predictive power and are proposed as a first step toward more efficient low-air-quality alarm systems.The procedure is described as adaptable to other ANN models for weather or geophysical forecasting; correlations across more monitoring stations are planned for future work.Source: §5 · p. 11