Session: Geographical Analysis, Urban Modeling, Spatial statistics A web-based autonomous weather monitoring system of the town of Palermo and its utilization for temperature nowcasting By: Giorgio Beccali, Maurizio Cellura, Simona Culotta, Valerio Lo Brano, Antonino Marvuglia UNIVERSITY OF PALERMO
Nowadays, the influence of microclimatic conditions on the energy behavior of buildings draws the attention of many researchers .
The enhanced urbanization, occurring both in developed and developing countries, led to the appearance of the Urban Heat Island (UHI) effect, notably known for the air temperature increase in the wide urban areas, compared to the conditions measured in the extra-urban areas .
An accurate analysis of the spatial and temporal evolution of the UHI needs a detailed collection of local weather data (recorded with a suitable sampling time) which can be accomplished only through an efficient weather monitoring system.
The Department of Energy and Environmental Researches (DREAM) of the University of Palermo has started to build up a network of weather stations displaced in different parts of the town.
Area: 6.25 Km 2 Area: 13.5 Km 2 Area: 45.6 Km 2 Installed in March 2008 July, 10 th 2007 MeteoPalermo1 March, 4 th 2008 Albunea May, 21 st 2007 Pizia February, 20 th 2007 Amaltea November, 30 th 2006 Cassandra November, 13 rd 2006 Morgana September, 26 th 2006 Merlino Installation date Weather station
Data acquisition and web publishing system The Linux server of DREAM connects to the shared folder of this PC where the file is stored and copies it into a local folder. The procedure is automated by a bash script and is repeated for each weather station. Re-formatting of the ASCII file through a Perl script A web server (Apache) is connected with the database server (MySQL) and an http service is available over TCP/IP network . Graphs and data digest : publicly available. Statistic elaborations and data download : protected by login and password. Every 30 minutes each weather station automatically generates an ASCII file containing the last 336 collected data and immediately transfers it (via GSM) to a MS Windows PC located at the DREAM building, in which a proprietary software is installed.
For temperature and humidity analysis interactive graphs have been created, that allow the user to select the desired reference period, to zoom in on a particular area of the chart or to change the scale of representation.
It is possible to display several polar diagrams showing the prevailing wind direction during the selected period. It is also possible to plot a polar diagram with the wind events subdivided in velocity classes (< 1 m/s; from 1 m/s to 2.5 m/s; > 2.5 m/s).
T he very short-term temperature forecasting (nowcasting) problem was treated as the identification of a dynamical system and tackled by the utilization of a Nonlinear Black-Box model based on a ANN.
The underlying hypothesis is that the system is signified by transfer function characterizations. As the behaviour of the system changes, the ANN model developed keeps track of the changes in the features and parameters of the system.
Thus, at any instant of time, it correctly simulates the given time-varying system, despite any significant change in its properties.
The analyses were accomplished by using Matlab 7.0:
NNSYSID Toolbox by M. Nørgaard ( Model selection );
System Identification Toolbox ( Parameters estimation and Model validation )
When widening the focus to also include identification of nonlinear dynamic systems, the problem of selecting model structures becomes more difficult.
By exploiting the typical capability of MultiLayer Perceptrons (MLP) to learn nonlinear relationships from a set of data, nonlinear extensions of the most common linear structures for time series prediction have been created.
By making this choice, the model structure selection is basically reduced to dealing with the following two issues:
Selecting the inputs to the network;
Selecting an internal network architecture
An often used approach is to reuse the input structures from the linear models substituting the internal architecture with a MLP network.
Nonlinear counterparts to the linear time series forecasting model structures are thus obtained by:
where: is the vector containing the adjustable parameters in the neural network known as weights ; g is the function realized by the neural network and it is assumed to have a feed-forward structure. is again the regression vector ;
Depending on the choice of the regression vector, different nonlinear model structures emerge. If the regression vector is selected as for ARX models, the model structure is called NNARX. Likewise, NNFIR, NNARMAX, NNOE and NNSSIF structures there exist .
The models were validated by observing the differences between the measured and the predicted temperatures for the week ranging from August 25 th to 31 th 2007, which had not been used for the training phase.
max min 0.39 1.27 Pizia 0.45 1.44 Morgana 0.68 2.11 MeteoPalermo1 0.35 1.13 Merlino 0.48 1.55 Cassandra 0.45 1.47 Amaltea MAE (°C) MAPE (%) Weather station
Results Week August 25 th to 31 th 2007 Morgana Cassandra Amaltea Pizia MeteoPalermo1 Merlino Output One-step ahead prediction
The developed web-based autonomous weather monitoring system was designed to study the spatial and temporal variation of the temperatures within the urban area of Palermo and investigate the influence of various factors on the UHI.
It represents a very powerful tool for urban monitoring.
In a future research activity a forecasting model with a wider time span will be implemented, in order to obtain the future evolution of the temperature with a relevant advance and use this information to study the time evolution of urban comfort conditions.
By using some of the existing outdoor comfort indices it will be possible to create dynamic maps of the actual and forecasted thermal hygrometric comfort conditions at urban scale and publish them on the website of DREAM.
Dr. Antonino Marvuglia e-mail: email@example.com tel: +39 091 236 139 web site: www.dream.unipa.it Thank you for your attention