| |
 |
UNIVERSITY OF BUCHAREST
FACULTY OF PHYSICS Guest
2024-11-22 1:30 |
 |
|
|
|
Conference: Bucharest University Faculty of Physics 2012 Meeting
Section: Physics and Technology of Alternative Renewable Energy Supplies
Title:
Weather Forecasting with Sensor Networks
Authors:
M. STANESCU
Affiliation:
University of Bucharest, Faculty of Physics, 3Nano-SAE Research Centre, ROMANIA
E-mail
stanescu.m.a@gmail.com
Keywords:
weather forecast, predict, temperature, sensor network, gaussian processes, time series
Abstract:
These days, the whole world concentrates on installing renewable energy resources. However, this endeavour holds lots of problems, the most serious being that renewable energy is greatly affected by weather conditions. The power produced is provided irregularly, raising issues concerning system stability and reliability. We work on a weather forecasting algorithm which may use any number of sensors in order to provide good weather estimations; these would allow accurately forecasts of the future power generation.
There have been many approaches to weather forecasting using data mining and Artificial Intelligence concepts, but they deal with measurements from only one sensor; the purpose of my project is to construct a model which uses records from a sensor network, to predict or recover weather data from any sensor within. Sensors that are close to one another, or in similar environments, will tend to make similar readings, a fact we could exploit with the right model. Such a model is feasible using the Gaussian Processes framework, which is able to represent and infer correlations and delays between different sensors.We have compared and contrasted a time series model which makes use of neural networks, a standard independent GP that treats sensors as independent and multi-output GPs that model the correlation between sensors. We found that the later are significantly more accurate, especially for the long term prediction, even when only a short time interval has been used for training.There are some additional benefits for the multi-output GP approach, including low computational requirements if using FITC or PITC approximations, and the possibility of interpolating between sensor readings to predict variables at missing sensors (sensors that have failed or are unavailable due to network outages). Results show that the algorithm is able to learn the correlations between sensors, and incorporate information from additional sensors to counterbalance the failures and missing data.
Note:This work was financially supported by contract PNCDI-II-IDEI/64/2011
|
|
|
|