Extended Abstract (268K)3.5
Mixing height short range forecasting through neural network modeling applied to radon and meteorological data
Antonello Pasini, CNR, Rome, Italy; and F. Ameli and M. Loré
Once recognized the complexity of the boundary layer physics and the difficulty in a mixing height short range forecasting if dealt with complete dynamics, here an approach to this prediction is introduced in terms of the joint application of neural network and box models. Due to the relevance of radon progeny measurements for an estimation of the diffusive properties of the atmospheric boundary layer, with particular reference to the estimation of the mixing height via a box model using these data, they are added to the most usual meteorological observations in order to obtain a more accurate characterization of the boundary layer state. In this paper we perform short range forecasts of the radon concentration and, through the further application of the cited box model, we obtain reliable predictions of the mixing height in nocturnal stable situations.
The neural model used is endowed with feed forward networks and backpropagation training, and was tested in the past on similar forecasting problems. Two strategies are adopted and compared: a time series approach using only inputs from time-delayed radon data or a synchronous pattern approach where meteorological and radon detections at a certain time t0 give in input the initial state of the system. In the time series approach good results are obtained through the application of a preprocessing method, in order to leave to the neural model the forecasting activity only on the hidden dynamics, that is to say on that part of the signal which is not reducible to known deterministic phenomena and periodicities. On the other hand, in the synchronous pattern approach the relevance of the different inputs is analyzed by means of bivariate statistical analyses and a network pruning is performed.
Session 3, Neural Networks
Tuesday, 11 February 2003, 8:30 AM-12:15 PM
Previous paper Next paper