A Comparison of Nonlinear Regression and Neural Network Models for Ground-Level Ozone Forecasting

ABSTRACT

A hybrid nonlinear regression (NLR) model and a neural network (NN) model, each designed to forecast next-day maximum 1-hr average ground-level O₃ concentrations in Louisville, KY, were compared for two O₃ seasons—1998 and 1999. The model predictions were compared for the forecast mode, using forecasted meteorological data as input, and for the hindcast mode, using observed meteorological data as input. The two models performed nearly the same in the forecast mode. For the two seasons combined, the mean absolute forecast error was 12.5 ppb for the NLR model and 12.3 ppb for the NN model. The detection rate of 120 ppb threshold exceedances was 42% for each model in the forecast mode. In the hindcast mode, the NLR model performed marginally better than the NN     

Accuracy and reliability of an automated air quality forecast system for ozone in seven Kentucky metropolitan areas

An automated forecast system for ozone in seven Kentucky metropolitan areas has been operational since 2004. The forecast system automatically downloads the required input data twice each day, produces next-day forecasts of metro area peak 8-h average ozone concentration using a computer coded hybrid nonlinear regression (NLR) model, and posts the results on a website. The automated models were similar to previous NLR models, first applied to forecasting ozone in the Louisville metro area. The forecast system operated reliably during the 2004 and 2005 O₃ seasons, producing at least one forecast per day better than 99% of the time. The forecast accuracy of the automated system was good. For all 2004 and 2005 forecasts, the mean absolute error was equal to 8.7 ppb, or 15.6% of the overall mean concentration. The overall detection rate of air quality standard exceedences was 56%, and the overall false alarm rate was 42%. In Louisville, the performance of the automated system was comparable to that of expert forecasters using the NLR model as a forecast tool.     

Fuzzy system models combined with nonlinear regression for daily ground-level ozone predictions

This study focuses on applying a Takagi–Sugeno fuzzy system and a nonlinear regression (NLR) model for ozone predictions in six Kentucky metropolitan areas. The fuzzy “c-means” clustering technique coupled with an optimal output predefuzzification approach (least square method) was used to train the Takagi–Sugeno fuzzy system. The fuzzy system was tuned by specifying the number of rules and the fuzziness factor. The NLR models were based in part on a previously reported, trajectory-based hybrid NLR model that has been used for years for forecasting ground-level ozone in Louisville, KY. The NLR models were each composed of an interactive nonlinear term and several linear terms. Using a common meteorological parameter set as input variables, the NLR models and the Takagi–Sugeno fuzzy systems model exhibited equivalent forecasting performance on test data from 2004. For all 2004 ozone season forecasts for the six metropolitan areas, the mean absolute error was 8.1 ppb for the NLR model and 8.0 ppb for the Takagi–Sugeno fuzzy model. When a nonlinear term (which was part of the NLR model) was included in the fuzzy model, the combined NLR–fuzzy model had slightly better performance than the original NLR model. For all 2004 metropolitan area forecasts, the mean absolute error of the NLR–fuzzy model forecasts was 7.7 ppb. These small differences may be statistically significant, but for practical purposes the performance of the fuzzy models was equivalent to that of the NLR models.     

An enhanced ozone forecasting model using air mass trajectory analysis

An enhanced ozone forecasting model using nonlinear regression and an air mass trajectory parameter has been developed and field tested. The model performed significantly better in predicting daily maximum 1-h ozone concentrations during a five-year model calibration period (1993–1997) than did a previously reported regression model. This was particularly true on the 28 “high ozone” days ([O₃]>120 ppb) during the period, for which the mean absolute error (MAE) improved from 21.7 to 12.1 ppb. On the 77 days meteorologically conducive to high ozone, the MAE improved from 12.2 to 9.1 ppb, and for all 580 calibration days the MAE improved from 9.5 to 8.35 ppb. The model was field-tested during the 1998 ozone season, and performed about as expected. Using actual meteorological data as input for the ozone predictions, the MAE for the season was 11.0 ppb. For the daily ozone forecasts, which used meteorological forecast data as input, the MAE was 13.4 ppb. The high ozone days were all anticipated by the ozone forecasters when the model was used for next day forecasts.     

An enhanced PM2.5 air quality forecast model based on nonlinear regression and back-trajectory concentrations

An enhanced PM_2.5 air quality forecast model based on nonlinear regression (NLR) and back-trajectory concentrations has been developed for use in the Louisville, Kentucky metropolitan area. The PM_2.5 air quality forecast model is designed for use in the warm season, from May through September, when PM_2.5 air quality is more likely to be critical for human health. The enhanced PM_2.5 model consists of a basic NLR model, developed for use with an automated air quality forecast system, and an additional parameter based on upwind PM_2.5 concentration, called PM24. The PM24 parameter is designed to be determined manually, by synthesizing backward air trajectory and regional air quality information to compute 24-h back-trajectory concentrations. The PM24 parameter may be used by air quality forecasters to adjust the forecast provided by the automated forecast system. In this study of the 2007 and 2008 forecast seasons, the enhanced model performed well using forecasted meteorological data and PM24 as input. The enhanced PM_2.5 model was compared with three alternative models, including the basic NLR model, the basic NLR model with a persistence parameter added, and the NLR model with persistence and PM24. The two models that included PM24 were of comparable accuracy. The two models incorporating back-trajectory concentrations had lower mean absolute errors and higher rates of detecting unhealthy PM2.5 concentrations compared to the other models.     

Evaluation of the Mesoscale Meteorological Model (MM5)-Community Multi-Scale Air Quality Model (CMAQ) performance in hindcast and forecast of ground-level ozone

This paper presents the first attempt to apply the Mesoscale Meteorological Model (MM5)-Community Multi-Scale Air Quality Model (CMAQ) model system to simulate ground-level ozone (O3) over the continental Southeast Asia (CSEA) region for both hindcast and forecast purposes. Hindcast simulation was done over the CSEA domain for two historical O3 episodes, January 26-29, 2004 (January episode, northeast monsoon) and March 24-26, 2004 (March episode, southwest monsoon). Experimental forecast was done for next-day hourly O3 during January 2006 over the central part of Thailand (CENTHAI). Available data from 20 ambient monitoring stations in Thailand and 3 stations in Ho Chi Minh City, Vietnam, were used for the episode analysis and for the model performance evaluation. The year 2000 anthropogenic emission inventory prepared by the Center for Global and Regional Environmental Research at the University of Iowa was projected to the simulation year on the basis of the regional average economic growth rate. Hourly emission in urban areas was prepared using ambient carbon monoxide concentration as a surrogate for the emission intensity. Biogenic emissions were estimated based on data from the Global Emissions Inventory Activity. Hindcast simulations (CSEA) were performed with 0.5 degree x 0.5 degree resolution, whereas forecast simulations (CENTHAI) were done with 0.1 degree x 0.1 degree hourly emission input data. MM5-CMAQ model system performance during the selected episodes satisfactorily met U.S. Environmental Protection Agency criteria for O3 for most simulated days. The experiment forecast for next-day hourly O3 in January 2006 yielded promising results. Modeled plumes of ozone in both hindcast and forecast cases agreed with the main wind fields and extended over considerable downwind distances from large urban areas.     

Norwegian Meteorological Institute’s real-time dispersion model snap (Severe Nuclear Accident Program): Runs for ETEX and ATMES II experiments with different meteorological input

The Norwegian Meteorological Institute (DNMI) has developed and implemented for operational use a real-time dispersion model Severe Nuclear Accident Program (SNAP) with capability for predicting concentrations and depositions of the radioactive debris from large accidental releases. SNAP has been closely linked to DNMI’s operational numerical weather prediction (NWP) models.How good are these predictions? Participation in ETEX has partly answered this question. DNMI used SNAP with LAM50S giving meteorological input for these real-time dispersion calculations. LAM50S Limited Area Model with 50 km grid squareswas DNMI’s operational NWP model in 1994 when ETEX took place.In this article we report on how SNAP performed in the first of the ETEX releases in near-real-time mode, using LAM50S—and in hindcast mode for ATMES II, using “ECMWF 1995: ETEX Data set (ATMES II)”as meteorological input data. These two input data sets came from NWP models with quite different characteristics but with similar resolution in time and space.The results from these dispersion simulations matched closely. Deviations early in the simulation period shrank to insignificant differences later on. Since both input data sets were based on “weather analysis” and had similar resolution in space and time, SNAP described the dispersion of the released material very similar in these two simulations.     

Trends in meteorologically adjusted ozone concentrations in six Kentucky metro areas, 1998-2002

Meteorologically adjusted ozone (O3) concentrations during five recent O3 seasons (1998-2002) were computed for six Kentucky metro areas using a nonlinear regression model originally developed for forecasting ground-level O3 concentrations. The meteorological adjustment procedure was based on modifying actual measured O3 concentrations according to model-predicted responses to climate departures with respect to a reference year. For all six Kentucky metro areas, meteorologically adjusted O3 concentrations declined over the five-year period. The linear best-fit rate of decline in mean adjusted O3 concentrations ranged from 0.9 to 2.6 ppb/yr for these metro areas; the average rate of decline was 1.6 ppb/yr. The rates of decline in meteorologically adjusted extreme value (e.g., 95th percentile) concentrations were approximately the same, but there is greater statistical uncertainty concerning the extreme value trends.     

Forecasting peak daily ozone levels: part 2--A regression with time series errors model having a principal component trigger to forecast 1999 and 2002 ozone levels

A modified time series approach, a Box-Jenkins regression with time series errors (RTSE) model plus a principal component (PC) trigger, has been developed to forecast peak daily 1-hr ozone (O3) in real time at the University of Wisconsin-Milwaukee North (UWM-N) during 1999 and 2002. The PC trigger acts as a predictor variable in the RTSE model. It tries to answer the question: will the next day be a possible high O3 day? To answer this question, three PC trigger rules were developed: (1) Hi-Low Checklist, (2) Discriminant Function Approach I, and (3) Discriminant Function Approach II. Also, a pure RTSE model without including the PC trigger and a persistence approach were tested for comparison. The RTSE model with DFA I successfully forecast the only two 1-hr federal exceedances (124 ppb), one in 1999 and one in 2002. In terms of the O3 100-ppb exceedances, 60-80% of the incorrect forecasts occurred with incorrect PC decisions. A few others may have been caused by unexpected O3-weather relations. When the three approaches used UWM-N data to forecast a 100-ppb exceedance somewhere in the Milwaukee, WI, metropolitan area, their performance was dramatically improved: the false alarm rate was reduced from 0.89 to 0.44, and the probability of detection was increased from 0.71 to 0.88.     

A performance evaluation of the National Air Quality Forecast Capability for the summer of 2007

This paper provides a performance evaluation of the real-time, CONUS-scale National Air Quality Forecast Capability (NAQFC) that supported, in part, its transition into operational status. This evaluation focuses primarily on discrete forecasts for the maximum 8-h O₃ concentrations covering the 4-month period, June through September, 2007, using measurements obtained from EPA's AIRNow network. Results indicate that the 2007 NAQFC performed as well or better than previous configurations, despite the expansion of the forecast domain into the western half of the nation that is dominated by complex terrain. The mean, domain-wide, season-long correlation was 0.70. When examined over time, the domain-wide correlations exhibit a fairly consistent nature, with values exceeding 0.60 (0.70) over 90% (55%) of the days. The NAQFC systematically over-predicted the 8-h O₃ concentrations, continuing a trend established by earlier NAQFC configurations, though to a lesser degree. The summer-long mean forecast value of 53.2 ppb was 4.2 ppb higher than the observed value, resulting in a domain-wide Normalized Mean Bias (NMB) of 8.7%. Most of the over-prediction is associated with observed concentrations less than 50 ppb. In fact the model tends to under-predict when concentrations exceed 70 ppb. As with the bias, the error associated with the latest configuration was also lower. The summer-long Root Mean Square Error of 13.0 ppb (Normalized Mean Error (NME) = 20.4%) represented marked improvements over earlier forecasts. Examination of the spatial distribution of both the NMB and NME reveals that the NAQFC was generally within 25% for the NME and 25% for the NMB over a majority of the domain. Several areas of poorer performance, where the NMB and NME often exceed 25% and in some cases 50%, were noted. These areas include southern California, where the NAQFC tended to under-predict concentrations (especially on weekends) and the southeast Atlantic and Gulf coasts regions, where the model over-predicted. Subsequent analysis revealed that the incorrect temporal allocation of precursor emissions was likely the source of the under-prediction in southern California, while inaccurate simulation of PBL heights likely contributed to the over-prediction in the coastal regions.     

Forecasting hourly PM<Subscript>10</Subscript> concentration in Cyprus through artificial neural networks and multiple regression models: implications to local environmental management

In the present work, two types of artificial neural network (NN) models using the multilayer perceptron (MLP) and the radial basis function (RBF) techniques, as well as a model based on principal component regression analysis (PCRA), are employed to forecast hourly PM₁₀ concentrations in four urban areas (Larnaca, Limassol, Nicosia and Paphos) in Cyprus. The model development is based on a variety of meteorological and pollutant parameters corresponding to the 2-year period between July 2006 and June 2008, and the model evaluation is achieved through the use of a series of well-established evaluation instruments and methodologies. The evaluation reveals that the MLP NN models display the best forecasting performance with R ² values ranging between 0.65 and 0.76, whereas the RBF NNs and the PCRA models reveal a rather weak performance with R ² values between 0.37-0.43 and 0.33-0.38, respectively. The derived MLP models are also used to forecast Saharan dust episodes with remarkable success (probability of detection ranging between 0.68 and 0.71). On the whole, the analysis shows that the models introduced here could provide local authorities with reliable and precise predictions and alarms about air quality if used on an operational basis.     

Multivariate Stochastic Models of Sulphur Dioxide Pollution in an Urban Area

Three multivariate stochastic mathematical models of daily SO₂ pollution in an urban area (Milan, Italy) during the heating season (mid-October/end of March) are illustrated in the paper. Each model is characterized by a different number of external inputs. Precisely, the first model has no inputs (it is simply an autoregressive relationship), the second one has a temperature input (roughly accounting for emission), the third one has two inputs (temperature and wind speed). From each model a real-time predictor is derived, namely a recursive relationship which, at the end of each day, allows future pollution levels to be forecast on the basis of current concentration and meteorological measurements. The quality of the forecast is rather satisfactory, even in episode situations. The improvements in forecast performance when turning from a predictor with less external inputs to a predictor with more external inputs (i.e., when exploiting more information about meteorology) are also pointed out in the paper.     

Application of artificial neural networks to modeling and prediction of ambient ozone concentrations

The deterministic modeling of ambient O3 concentrations is difficult because of the complexity of the atmospheric system in terms of the number of chemical species; the availability of accurate, time-resolved emissions data; and the required rate constants. However, other complex systems have been successfully approximated using artificial neural networks (ANNs). In this paper, ANNs are used to model and predict ambient O3 concentrations based on a limited number of measured hydrocarbon species, NOx compounds, temperature, and radiant energy. In order to examine the utility of these approaches, data from the Coastal Oxidant Assessment for Southeast Texas (COAST) program in Houston, TX, have been used. In this study, 53 hydrocarbon compounds, along with O3, nitrogen oxides, and meteorological data were continuously measured during summer 1993. Steady-state ANN models were developed to examine the ability of these models to predict current O3 concentrations from measured VOC and NOx concentrations. To predict the future concentrations of O3, dynamic models were also explored and were used for extraction of chemical information such as reactivity estimations for the VOC species. The steady-state model produced an approximation of O3 data and demonstrated the functional relationship between O3 and VOC-NOx concentrations. The dynamic models were able to the adequately predict the O3 concentration and behavior of VOC-NOx-O3 system a number of hourly intervals into the future. For 3 hr into the future, O3 concentration could be predicted with a root-mean squared error (RMSE) of 8.21 ppb. Extending the models further in time led to an RMSE of 11.46 ppb for 5-hr-ahead values. This prediction capability could be useful in determining when control actions are needed to maintain measured concentrations within acceptable value ranges.     

BP神经网络对蚯蚓滤池处理COD的模拟预测

基于蚯蚓滤池处理去除污染物的非线性特点,利用BP神经网络建立了蚯蚓滤池处理COD的基本模型结构。同时对实验数据进行了验证和预测,通过权值贡献率分析确定了各种输入因素对COD出水浓度的影响。结果表明:COD的出水模型预测值与实际值平均误差较小,模型稳定,预测效果好。输入神经元为4,隐含神经元为8,输出神经元为1,学习速率为0.1,动量为0.1,训练次数为10 000的BP神经网络模型,预测的COD出水值最接近真实值。COD进水浓度对COD出水影响最大,符合理论研究结果。BP神经网络模型建立的成功为后续生活污水智能化控制的研究提供了相应的理论基础。     

Ensemble-based data assimilation and targeted observation of a chemical tracer in a sea breeze model

We study the use of ensemble-based Kalman filtering of chemical observations for constraining forecast uncertainties and for selecting targeted observations. Using a coupled model of two-dimensional sea breeze dynamics and chemical tracer transport, we perform three numerical experiments. First, we investigate the chemical tracer forecast uncertainties associated with meteorological initial condition and forcing error. We find that the ensemble variance and error builds during the transition between land and sea breeze phases of the circulation. Second, we investigate the effects on the forecast variance and error of assimilating tracer concentration observations extracted from a truth simulation for a network of surface locations. We find that assimilation reduces the variance and error in both the observed variable (chemical tracer concentrations) and unobserved meteorological variables (vorticity and buoyancy). Finally, we investigate the potential value to the forecast of targeted observations. We calculate an observation impact factor that maximizes the total decrease in model uncertainty summed over all state variables. We find that locations of optimal targeted observations remain similar before and after assimilation of regular network observations.     

An Urban Diffusion Simulation Model For Carbon Monoxide

A relatively simple Gaussian-type diffusion simulation model for calculating urban carbon monoxide (CO) concentrations as a function of local meteorology and the distribution of traffic is described. The model can be used in two ways: (1) in the synoptic mode, in which hourly concentrations at one or many receptor points are calculated from historical or forecast traffic and meteorological data; and (2) in the climatological mode, in which concentration frequency distributions are calculated on the basis of long-term sequences of input data. For model evaluation purposes, an extensive field study involving meteorological and air-quality measurements was conducted during November-December 1970 in San Jose, Calif., which has an automated network to provide traffic data throughout the central business district. Model refinements made on the basis of the data from this experimental program include the addition of a street-canyon submodel to compensate for the important aerodynamic effects of buildings on CO concentrations at streetside receptors. The magnitude of these effects was underscored by the concentrations measured on opposite sides of the street in San Jose, which frequently differed by a factor of two or more. Evaluation of the revised model has shown that calculated and observed concentration frequency distributions for street-canyon sites are in good agreement. Hour-average predictions are well correlated with observations (correlation coefficient of about 0.6 to 0.7), and about 80 percent of the calculated values are within 3 ppm of the observed hour-average concentrations, which ranged as high as 16 ppm.     

Ozone and grosswetterlagen

Meteorological conditions have a decisive impact on surface ozone concentrations. In this study, an empirical model is used to explain the interdependence of ozone and grosswetterlagen. Different meteorological parameters such as air temperature, global solar radiation, relative humidity, wind direction and wind speed are used. Additional nitric oxide (NO) was taken as a representative for the emission situation and ozone maximum of the preceding day in order to evaluate the development of the photochemical situation. The dataset includes data collected over a period of three years (1992–1994) from three stations outside of Munich and one in the center of Munich. All values become variables by calculating means, sums or maxima of the basic dataset consisting of half-hour means. Seasonal periodicity of data is detected with Fourier analysis and eliminated by a division method after computing a seasonal index. The dataset is divided into three different grosswetterlagen groups, depending on main wind direction. One mostly cyclonic (westerly winds), one mixed (alternating winds) and one only anticyclonic (easterly winds). The last is completed with one summertime group including values from April to August. Factor analysis is performed for each group to obtain independent linear variable combinations. Overall, relative humidity is the dominant parameter, a typical value indicating meteorological conditions during a grosswetterlage. Linear multiple regression analysis is performed using the factors obtained to reveal how the ozone concentrations are explained in terms of meteorological parameters and NO. The results improve from cyclonic to anticyclonic grosswetterlagen in conformance with the increasing significance of photochemistry, indicated by the high solar radiation and high temperature, and the low relative humidity and low wind speed. The explained variance r² reaches its maximum with more than 50 % of the time in Munich center. This empirical model is applicable to the forecasting of local ozone maximum concentrations with a total standard error deviation of 8.5 to 12.8 % and, if ozone concentrations exceed 80 ppb, with a standard error deviation of 5.4 to 9.5 %.     

Ozone and grosswetterlagen

Meteorological conditions have a decisive impact on surface ozone concentrations. In this study, an empirical model is used to explain the interdependence of ozone and grosswetterlagen. Different meteorological parameters such as air temperature, global solar radiation, relative humidity, wind direction and wind speed are used. Additional nitric oxide (NO) was taken as a representative for the emission situation and ozone maximum of the preceding day in order to evaluate the development of the photochemical situation. The dataset includes data collected over a period of three years (1992–1994) from three stations outside of Munich and one in the center of Munich. All values become variables by calculating means, sums or maxima of the basic dataset consisting of half-hour means. Seasonal periodicity of data is detected with Fourier analysis and eliminated by a division method after computing a seasonal index. The dataset is divided into three different grosswetterlagen groups, depending on main wind direction. One mostly cyclonic (westerly winds), onemixed (alternating winds) and one onlyanticyclonic (easterly winds). The last is completed with one summertime group including values from April to August. Factor analysis is performed for each group to obtain independent linear variable combinations. Overall, relative humidity is the dominant parameter, a typical value indicating meteorological conditions during a grosswetterlage. Linear multiple regression analysis is performed using the factors obtained to reveal how the ozone concentrations are explained in terms of meteorological parameters and NO. The results improve from cyclonic to anticyclonic grosswetterlagen in conformance with the increasing significance of photochemistry, indicated by the high solar radiation and high temperature, and the low relative humidity and low wind speed. The explained variance r² reaches its maximum with more than 50 % of the time in Munich center. This empirical model is applicable to the forecasting of local ozone maximum concentrations with a total standard error deviation of 8.5 to 12.8 % and, if ozone concentrations exceed 80 ppb, with a standard error deviation of 5.4 to 9.5 %.     

Prediction of the vertical profile of ozone based on ground-level ozone observations and cloud cover

A number of statistical techniques have been used to develop models to predict high-elevation ozone (O3) concentrations for each discrete hour of day as a function of elevation based on ground-level O3 observations. The analyses evaluated the effect of exclusion/inclusion of cloud cover as a variable. It was found that a simple model, using the current maximum ground-level O3 concentration and no effect of cloud cover provided a reasonable prediction of the vertical profile of O3, based on data analyzed from O3 sites located in North Carolina and Tennessee. The simple model provided an approach that estimates the concentration of O3 as a function of elevation (up to 1800 m) based on the statistical results with a +/- 13.5 ppb prediction error, an R2 of 0.56, and an index of agreement, d1, of 0.66. The inclusion of cloud cover resulted in a slight improvement in the model over the simple regression model. The developed models, which consist of two matrices of 24 equations (one for each hour of day for clear to partly cloudy conditions and one for cloudy conditions), can be used to estimate the vertical O3 profile based on the inputs of the current day's 1-hr maximum ground-level O3 concentration and the level of cloud cover.