92nd American Meteorological Society Annual Meeting (January 22-26, 2012)

Wednesday, 25 January 2012: 8:45 AM
Optmisation of the AIR TURBULENCE PARAMETERS by Neural Network MODELING
Room 242 (New Orleans Convention Center )
Armando Pelliccioni, ISPESL, Monteporzio Catone, Italy; and C. Gariazzo

One of more important questions in the simulation of the air turbulence parameters by Neural Network is the choice of the right variables to be used in the simulation and their underlined importance in the phenomena to be simulated. What process represents real data measurements and which kind of information do they take into account? The observed data are strictly linked with the natural solution of the transport equation, the turbulence conditions and the thermal exchange between the ground and upper layer of the air level. In using the NN to reproduce the turbulence observed data, it is necessary to assume some relevance to one important question. The goal of intelligent methods is to find the “true” deterministic solution of a certain phenomena by using the information available in an experimental dataset which contains observations of all processes influencing it. In order to do this for air turbulence in the planetary boundary layer, the performances of a NN have been tested using different input variables to reproduce the main turbulence parameters such as turbulent kinetic energy (TKE), Monin-Obukhov length (L) and friction velocity (u*). The idea concerns with each variable contain specific information linked with the simulated process. Although NN can be considered as a universal estimator of any function, in the case of turbulence simulation care has to be paid in the selection of information during both the training and generalization phases of NN parameters. These aspect can play a negative sense when the generalization data never can be sight by the NN itself during the training phase. In particular the future data could not contain the information embedded in data used during the training phase, so the best NN have to analyze the information inside data before the analysis. In this work it will be introduced the concept of the Scientific Information Theory (SIT) where all direct and indirect relations with the available variables are used in the simulation. This concept is fundamental for the right choice of the variables to be used during the NN simulation and the information is structured in different levels. It is fundamental to understand that different information levels can influence the NN model in a very significant way. More information is transported by the variables, easier and accurate is the task made by the intelligent models to reproduce the phenomena. To demonstrate what described above, a test using different input data to reproduce the TKE, L and U* has been carried out. A large data set of meteorological measurements has been collected, during the 2011 field campaigns in a city park in the centre of the Rome area, using two high frequency sonic anemometers, sensors of absolute and differential temperature, and other equipments to collect conventional meteorological data. Usually turbulence parameters such as the Monin-Obukhov length, the turbulent kinetic energy and the friction velocity can be computed by meteorological measurements using sonic anemometers or by means of the surface wind and temperature gradients measurements. As input data to the NN, the following variables have been considered: the standard deviations of vertical and horizontal wind, the temperature, wind and pressure at some level, the surface gradients of wind and temperature. The NN is tested using different combination of input variables, starting from the more simple data such as could the Pasquill stability classes, up to the most complex data linked with the direct turbulence measurements. To demonstrate the importance of the information inside input data selection, the simulations using some exogenous variables are given (as the hour of day and Julian day). As NN architecture have been used the Multilayer Perceptron, with different choice of the hidden units. The hidden unit depends by the complexity task to achieve to simulate the target variables. To reproduce the TKE and U* the number are relatively small (from 4 to 8 hidden neurons), while for the L the number goes from 8 to 12 neurons. These are in according with the well known facts that the Monin-Obukhov length is more difficult to simulate respect to the direct variables link to the physics of the system. Also the performances of NN to reproduce L, TKE and U are quite different. Results are referred to the generalization data set. The observed correlation for the last two variables is very good, showing the values of 0,83 up to 0,92, depending by the choice of the activation function and by the algorithm used for the minimization of the total errors. At the contrary, the correlation observed for L is always bad (from to 0,32 to 0,55). The bad behavior of NN to reproduce L is due to the facts that L is calculated as the ratio between mechanical and thermal fluxes, so that some numerical divergence occurs when the estimate of the thermal fluxes is close to the zero values (during the transition day-night and vice versa). To eliminate this bad behavior of NN, it has been test one alternative strategy to simulate L. In ours simulation it was calculated the simulation of L using the combination of all physics variables that intervene in the calculation of L itself. So doing, the performance of NN is increase and the indirect calculation of L improves in meaningful ways the correlation (from 0,65 to 0,85). The work demonstrates that the optimization on NN to simulate the turbulence variables by simple meteorological measurements is possible if the direct variables connect with the physics of the system is considered as target variables. The Monin-Obukhov length confirms that is very hard to simulate it directly by NN and a good estimate can be only obtained considering the indirect estimation of L. In according with the concepts associate to the SIT, the choice of the input variables are strictly link with the information associate to every ones and the Neural Net models results more complex when the variables used to simulate the physics are far from the direct information link to the system.

Supplementary URL: