PaperlayerView exact revision on arXiv

Source-linked AI summary

A 24-h forecast of ozone peaks and exceedance levels using neural classifiers and weather predictions

A. Dutot, Joseph Rynkiewicz, F. Steiner, J. Rude

arXiv:0802.3969v1stat.APmath.STstat.TH

TL;DR

Real-time forecasts of ozone peaks can help air-quality agencies meet forecasting and communication responsibilities. This paper combines a neural network with a classifier, using weather predictions, and reports a 78% Success Index with six of seven exceedances detected during validation.

  • Problem

    The paper addresses ozone-peak forecasting, which air-quality agencies need to inform authorities and the public and assess emission-reduction impacts.

    Source: 1) Introduction · p. 4

  • Method

    The method combines an MLP-based neural model with a neural classifier and uses French National Weather Service model output statistics as predictors.

    Source: 2-3) Statistical model · p. 7

  • Results

    78% Success Index and 8% False Alarm Rate were reported for the classifier, which detected 6 of 7 exceedances during validation.

    Source: 4) Conclusion · p. 21

  • Takeaways & Limitations

    The authors report the classifier's 78% Success Index and 8% False Alarm Rate as the best combined performance among the models.

    Source: 4) Conclusion · p. 21

    Solar radiation and upper-air ventilation data were unavailable for the model inputs.

    Source: 2-2) Data set · p. 7

Abstract

from arXiv · show

A neural network combined to a neural classifier is used in a real time forecasting of hourly maximum ozone in the centre of France, in an urban atmosphere. This neural model is based on the MultiLayer Perceptron (MLP) structure. The inputs of the statistical network are model output statistics of the weather predictions from the French National Weather Service. With this neural classifier, the Success Index of forecasting is 78% whereas it is from 65% to 72% with the classical MLPs. During the validation phase, in the Summer of 2003, six ozone peaks above the threshold were detected. They actually were seven.

1) Introduction

Air-quality agencies need forecasts of pollution peaks to inform the public and authorities and assess emission reductions. Statistical models are generally preferred for real-time forecasting because three-dimensional air-quality models are time-consuming and require large initialization databases.

1) Introduction

  • Air-quality agencies are tasked with forecasting pollution peaks, informing authorities and the public, and assessing emission-reduction impacts.These responsibilities motivate the paper's focus on air-quality forecasting.Source: 1) Introduction · p. 4
  • Statistical models are generally preferred for real-time forecasts over three-dimensional models because the latter are time-consuming and need very large databases to initialize and run.The paper distinguishes models integrating chemistry and transport from statistical models that connect meteorological conditions to pollutant levels.Source: 1) Introduction · p. 4
  • For ozone, daily weather variation best explains day-to-day level variability, and persistent summer conditions conducive to photochemical smog lead to ozone episodes.The study uses meteorological parameters as predictors in a statistical model.Source: 1) Introduction · p. 4

2) Method

The study develops an ozone-peak forecasting method using weather forecasts and a neural network coupled to a classifier. It uses an MLP-based regression model, pruning for regularisation, and leverage-based prediction intervals.

2) Method

  • The method uses French Weather Service model output statistics as readily available predictors for ozone-peak forecasts.The forecast input data are intended to support an operational ozone forecasting system.Source: 2-1) Strategy · p. 5
  • Prior ozone regression approaches are described as limited in modelling extreme values, motivating the test of a neural network coupled with a neural classifier for threshold exceedance prediction.The paper also notes that nonlinear neural methods have provided better representation of extreme values than linear methods.Source: 2-1) Strategy · p. 5
  • The dataset covers three urban monitoring stations in Orléans over 1999–2003, with April–September data used for the spring–summer model and validity also tested year-round.Ozone forecasts target daily maxima of hourly mean concentrations.Source: 2-2) Data set · pp. 5–6
  • The model inputs comprise 74 meteorological variables and ozone persistence, derived from weather-bulletin parameters predicted at differing time intervals.The variables were chosen because they were available in the forecast bulletin; solar radiation and upper-air ventilation data were unavailable.Source: 2-2) Data set · p. 6
  • The statistical model is an autoregressive neural model with exogenous inputs (NARX), implemented as a single-hidden-layer MLP regression.The MLP represents a nonlinear relationship between predictor inputs and the estimated output.Source: 2-3) Statistical model · p. 7
  • Poor local minima and overtraining are identified as MLP pitfalls; random weight initialisations are used to address the former.The authors note that multiple random initialisations are time-consuming but can be parallelised.Source: 2-3) Statistical model · p. 8
  • Overtraining is addressed by stepwise pruning that minimizes a BIC-like criterion, eliminating irrelevant weights and potentially inputs or hidden units.The procedure starts with all possible inputs and excess hidden neurons, then removes units while the criterion decreases; unlike early stopping, it allows learning and validation data to be combined into a larger learning set.Source: 2-3-1) Regularisation scheme · pp. 8–9
  • Prediction intervals are computed using training-example leverage, under full-rank and asymptotic conditions.The cited approach uses the model's residual standard deviation in the interval calculation.Source: 2-3-2) Confidence interval of prediction · pp. 9–10

3) Results

The results report the neural classifier’s performance on Summer 2003 validation data, including its Success Index and False Alarm Rate at station 3 and scores at two additional stations. They also describe comparisons with the logistic model and MLP global-fit performance, as well as the classifier’s sigmoid-output design.

3-3) Results of the statistical evaluation of the model performance

3) Results

  • Stepwise selection retained 8 input parameters, and the BIC optimum used 1 hidden-layer neuron.The network design was chosen at station 3, Saint Jean de Bray, and its architecture was reproduced at Préfecture and La Source.Source: 3-1) Inputs selection · p. 11

3-3-1) Evaluation of the global fit agreement

  • 92% agreement was achieved by the three MLPs, with MLP1 most accurate at RMSE unsystematic 12 µg m-3 and MAE 15 µg m-3.MLP1 had the lowest mean bias error; MLP1, MLP2 and MLP3 biases were 5, 11 and 7 µg m-3, while CHIMERE’s was −13 µg m-3.Source: 3-3-1) Evaluation of the global fit… · pp. 15–16

3-3-2) Evaluation of the exceedance indexes

  • MLP3 achieved 72% SI with a 14% FAR, while MLP1 had 65% SI and the lowest FAR, 6%.MLP3 produced 12 false alarms against 7 observed exceedances; MLP1’s SI was about the same as pure persistence.Source: 3-3-2) Evaluation of the exceedance… · p. 16

3-3-3) Development of a neural classifier to perform the exceedance levels

  • The classifier replaced MLP1’s identity output function with a sigmoid while retaining the same neural structure.Its output was interpreted as the probability of exceeding 180 µg m-3, with outputs of at least 0.5 classified as exceedances; it used the same predictors, training data and validation set.Source: 3-3-3) Development of a neural… · pp. 16–17
  • 78% SI and 8% FAR were achieved by the classifier, versus 72% SI and 14% FAR for MLP3.The classifier detected 6 of 7 observed ozone peaks and made 6 false predictions; MLP3 correctly predicted 4 of 7 exceedances.Source: 3-3-3) Development of a neural… · p. 18
  • Model p-value 0.0001, residual p-value 1.00, and chi-square goodness-of-fit p-value 0.76 supported a significant relationship and adequate logistic fit.The residual result indicated the model was not significantly worse than the best possible model for the dataset, and the goodness-of-fit test gave no reason to reject adequacy.Source: 3-3-3) Development of a neural… · pp. 18–19
  • 78% Success Index for the neural classifier exceeded 66% for logistic regression, which detected all observed exceedances but generated 29 false alarms.The logistic model also correctly forecast the isolated peak at the end of September that the classifier missed.Source: 3-3-3) Development of a neural… · p. 19

3-3-4) Forecasting on the other two stations

  • Using MLP1’s same predictors and hidden layer, retrained for each station on 1999–2002 and validated in 2003, the three station models had similar regression slopes and acceptable residuals.Standardized residual variation was less than ±2, with no special pattern; the models were therefore globally validated for fit agreement.Source: 3-3-4) Forecasting on the other two… · pp. 19–20
  • 89% Success Index at station 1 accompanied correct forecasts of all observed exceedances, while station 2 had 6 of 8 exceedances forecast and a 70% Success Index.Source: 3-3-4) Forecasting on the other two… · p. 20

4) Conclusion

The neural classifier gives the best combined exceedance-forecasting performance in the 2003 validation, while the unbalanced MLP gives the best global-fit scores. NEUROZONE was implemented in real time in the Orléans region for 24-hour-ahead forecasts.

4) Conclusion

  • The unbalanced MLP had the best 24-hour global-fit scores, with RMSE unsystematic = 12 µg m-3 and MAE = 15 µg m-3.The validation compared one MLP and two pattern-balanced MLPs with deterministic and persistence references over April–September 2003.Source: 4) Conclusion · p. 20
  • The pattern-balanced MLP reached a 72% Success Index but had the highest False Alarm Rate, at 14%.This difficulty in choosing an exceedance model motivated development of the neural classifier.Source: 4) Conclusion · p. 20
  • 78% Success Index, 8% False Alarm Rate, and 6 of 7 exceedances found make the neural classifier the best combined performer.Its output is interpreted as the probability of exceeding the 180 µg m-3 threshold.Source: 4) Conclusion · pp. 20–21
  • NEUROZONE was implemented in real time in the Orléans region, with forecasts focused on a 24-hour lead time.The system is retrained each year using the previous validation data; longer lead times are a possible future extension.Source: 4) Conclusion · pp. 20–21

Figures and Tables captions

The captions identify figures and tables covering ozone time series, network selection and structure, forecast comparisons, exceedance probabilities, residuals, and dataset and model-performance measures.

Figures and Tables captions

  • Figure 1 tracks hourly maximum ozone during summer from 1999 to 2003.The caption specifies the summer period as April to September.Source: Figure 1 · p. 27
  • Figures 2 and 3 show validation-phase BIC-like criterion by hidden-layer size and the network structure after REGRESS optimization.The zero-neuron MLP is identified as a multi-linear regression.Source: Figure 2 · p. 28, Figure 3 · p. 29
  • Figure 4 compares MLP1 and deterministic-model time series during validation, with an August 2003 zoom.The caption identifies the models by red and green, respectively.Source: Figure 4 · p. 30
  • Figure 5 plots observed and classifier-forecast probabilities of exceeding the European threshold during April–September 2003.Observations are black and classifier forecasts red; the caption also identifies logistic-model forecasts as blue squares.Source: Figure 5 · p. 31
  • Figures 6 and 7 present measured-versus-forecast maximum ozone concentrations at three stations and standardized residuals on validation data.The Figure 6 caption is truncated after identifying the three-station scatter plot.Source: Figure 6 · p. 32, Figure 7 · p. 33
  • Tables 1–3 cover training and validation set composition, validation performance measures for all models, and exceedance-forecasting measures for three stations.Table 1 caption says bracketed entries give numbers of values; the performance-table captions do not provide cell values here.Source: Table 1 · p. 34, Table 2 · p. 35, Table 3 · p. 36
Loading 0802.3969v1…