Neural Network Models for Air Quality Prediction: A Comparative Study S V Barai A.K.Dikshit Sameer Sharma Department of Civil Engineering IIT Kharagpur
Presentation Outline ……. � Background on Neural Forecaster � Objectives � Developed Neural Air Quality Predictors � Data Collection and Data Analysis � Performance Evaluation � Results and Discussion � Future Projections � Closing Remarks
Generic Air Quality Forecaster Air Quality Input Variables Parameter Equation or Mapping Mathematical Model Model Output Air Quality Forecasts
Neural Air Quality Forecaster Air Quality Input Variables Parameter Equation or Mapping Mathematical Model Model Neural Output Air Quality Networks Forecasts
Objectives � Data Collection -multiple air quality parameters - Containing daily average pollutant concentrations at a specific location. � Implementation of variations of neural network models for predicting air quality. � Identification of suitable air quality prediction model(s) for hourly (short-term) data and Yearly (long-term) data. � Exhaustive simulation Study using various models. � Performance study of models.
Implementation of Neural Models � Recurrent Network Model (RNM) � Change Point Detection Model (CPDM_RNM) � Sequential Network Construction Model (SNCM) � Self Organizing Feature Maps Model (SOFM)
Data Collection and Analysis
Case Study 1- Long Term � Seven parameters namely VOC (volatile organic carbon), NOX (oxides of nitrogen), CO (carbon monoxide), SO2 (sulphur dioxide), PM10 (particulate matter with size less than 10 microns), PM2.5 (particulate matter with size less than 2.5 microns) and NH3 (ammonia). � All concentrations are in micrograms per cubic meter. � Annual average data for 15 years from 1985 to 1999 � Source:US EPA website www.epa.gov. � Data for 115 counties of California State in USA has been collected
Case Study 2- Short Term � Three parameters namely RPMA (Respiratory Particulate Matter Average), SO 2 (sulphur dioxide) and NO 2 (nitrogen dioxide) is collected for Delhi State at nine locations. � Daily average concentrations for last two years from 3/7/2000 to 20/8/2001. � Tata Energy Research Institute web site www.teri.in � Data for Ashram Chowk has been used for carrying out simulation
Statistical Properties Case Study 1
Statistical Properties Case Study 2
Model Performance Evaluation
Error Evaluation Percentage Error = (target – output) / target *100
Results and Discussion
Case Study 1 – Long Term
RNM Parameters
CPDM_RNM Parameters
SNCM Parameters
SOFM Model Parameters
Models Performance
VOC emissions - SOFM Case Study 1 80 70 VOC (micg/m3) 60 50 Target 40 Output 30 20 10 0 1994 1995 1996 1997 1998 1999 Year
SO 2 emissions - SOFM Case Study 1 300 250 SO2 (micg/m3) 200 Target 150 Output 100 50 0 1994 1995 1996 1997 1998 1999 Year
PM10 emissions - SOFM Case Study 1 70 60 PM10 (micg/m3) 50 40 Target 30 Output 20 10 0 1994 1995 1996 1997 1998 1999 Year
Observations � The models in general could predict with some accuracy. � Self-organizing Feature Map (SOFM) based model has performed extremely well in comparison to other models. � Typical results of SOFM for VOC emissions, CO emissions, and NOx emissions results high-light the very good performance of the model. The discrepancy observed in the model prediction can be due to the modeling of the problem.
Observations (Contd.) � Demonstrated an example of an annual average emission (long-term) data prediction using various neural networks models for a very limited dataset. Models in general have performed reasonably well even with the limited historical data. It is expected that availability of more annual average emission data can improve the performance of models studied.
Case Study 2 – Short Term
RNM Parameters
CPDM_RNM Parameters
SNCM Parameters
SOFM Parameters
Models Performance
RPMA emissions - SOFM Case Study 2 350 300 RPMA (micg/m3) 250 200 Target 150 Output 100 50 0 318 322 325 329 332 336 339 350 378 381 Sample Number
SO 2 emissions - SOFM Case Study 2 18 16 14 SO2 (micg/m3) 12 Target 10 8 Output 6 4 2 0 325 332 336 339 343 346 360 364 378 381 Sample Number
NO 2 emissions - SOFM Case Study 2 120 100 NO2 (micg/m3) 80 Target 60 Output 40 20 0 339 350 353 357 360 364 367 371 374 381 Sample Number
Observations � Models could predict with modest accuracy. � SOFM based model has performed extremely well in comparison to other models. � Typical results of SOFM for RPMA emissions, SO2 emissions, and NO2 emissions showed reasonably good match between model predictions with target prediction.
Future Projections � Models can have as inputs data from multiple sources, such as historical air quality measurements, meteorological data etc. � Models can have along with emissions data, episode levels definition, historical measurements of surface and upper air meteorological data. � Models should be able to give predictions for the following three different time windows: 1 day, 1 week and 1 month predictions.
Future Projections (Contd.) � Models should have an easy to use interface and should be able to pre-sent the results in an understandable way to non-computer experts. � The model parameters and architecture of models in this project were arrived with trial and error. One can arrive at optimal and better performance model after carrying out systematic studies on networks models and their parameters using optimization techniques such as Genetic
Closing Remarks � Study was carried out on air quality forecasting using various neural network models: RNM, CPDM_RNM, SNCM and SOFM. � The study was focused at preliminary investigation of single variable based time series prediction. � The investigation was carried out for long-term as well as short-term air quality data set. � Self-Organizing Features Maps (SOFM) used for time series prediction came up as the best tool for time series forecasting.
Closing Remarks (Contd.) � These were found to be very useful for large training datasets. � The results shown here are indications that the neural network techniques can be useful tool in the hands of practitioners of air quality management and prediction. In that case, practitioners need not know even about the development of the model. � The models studied in this study are easily implemented, and they can deliver prediction in real time, unlike other modeling techniques. The models can very well easily deal with input noise and uncertainty
Pollution Free Solution
Acknowledgement Images used in the presentations are obtained from various web-sources. All these web-sources are gratefully acknowledged.
Recommend
More recommend
