Extended Abstract (332K)J20.5
EnviNNet: a Neural Network for hindcasting PM10 in Urban Phoenix
M. C. Mammarella, ENEA, Rome, Italy; and G. Grandoni, P. Fedele, R. A. DiMarco, H. J. S. Fernando, and P. Hyde
The Phoenix metropolitan area is currently designated as 'serious' with regard to violation of the U.S. National Ambient Air Quality Standards (NAAQS) for particulate matter of aerodynamic diameter less than 10µm (PM10). Most of the severe PM10 violations therein have been attributed to regional natural exceptional events or local exceptional (episodic) events associated with windblown dust emanating from area sources such as construction and agricultural sites, vacant lots and alluvial channels. In general, deterministic models have been used to predict these episodes, but the unavailability of recent pollution inventories and the very complexity of models have been the bane in developing operational forecasting tools based on deterministic models. To this end, ENEA has been working with ASU to develop a stochastic model based on neural networks, which is called EnviNNet. This presentation deals with the results of such attempts.
The neural networks (NN) had already been used in air quality forecasting with considerable success. In developing EnviNNet, the selection of a subset of input variable needed to be done with care, paying attention to site-specific exceptional events, including time lag effects. The data noise needed to be considered in order to adapt satisfactorily the non-linear dynamic interaction between meteorological and pollution related processes. The EnviNNet employs 3-layer MLP network architecture with hidden nodes, full connection between layers and no connection between neurons in the same layer topology and exponential transfer function. In training the network, characteristics of high, low and episodic air pollution events at a particular data site were taken in the account. Since PM10 concentrations are heterogeneous throughout the year, to ensure that EnviNNet is robust and adaptive to the local climate, specific input data subsets were built by combining time-section data from different years. The meteorological and pollution data were taken from selected four-to-six-month windows as well as time periods that show noteworthy patterns. EnviNNet has shown superior performance compared to conventional deterministic models in predicting PM10 peaks. Efforts are underway to extend the EnviNNet to other data stations in the area and in the future to set up a PM10 prediction system for the Phoenix area based on a network of new NN stations.
Joint Session 20, Air Quality in Cities—Criteria Pollutants (Joint with the Meteorological Aspects of Air Pollution Committee)
Thursday, 15 January 2009, 8:30 AM-9:45 AM, Room 124A
Previous paper