The importance of being able to accurately predict weather cannot be overstated; beyond predicting catastrophic events such as storms or unusually severe spells of cold or heat, weather has a dramatic effect on a population’s day to day health and behaviour and has a significant impact on the economy.
Trying to model the behaviour of an extremely dynamic weather system such as the UKs is very challenging. Discovering which parameters are significant and what are the ramifications of specific changes in weather data patterns is an extremely complex problem to solve.
This complexity has led a number of researchers to stop trying to model the complex physical mechanics and interactions of the weather and instead to use a “brute force” approach to see if the data alone may be used. In other words to see if the massive quantity of data generated by the thousands of weather stations may be successfully utilised by a neural network in order to predict weather. This is generally recognised as one of the early uses of Big Data for analytical purposes.
Historically researchers have used relatively traditional neural networks to try to predict specific weather characteristics such as rainfall. Some researchers have used more exotic networks or have used an ensemble of neural networks in an effort to solve this complex problem.
To date few researchers have used newer generations of neural networks/AI algorithms for this purpose, which continues to make this an open area for exploration.
For the competition, a localised weather dataset (Oak Park weather station) was taken from closest weather station to the IJCNN2015 conference venue. Please click the link below to download the Dataset.
The dataset covers historic daily values for the period between 1st January 2007 and 30th September 2014, for the following weather measurements:
Minimum Temperature and Maximum Temperature
(Rainfall data is included in case they are needed to contribute to enhancing the accuracy of the prediction)
Any type and/or combination of neural networks can be used to make the following predictions:
The generalisation dataset should be used to test the neural networks performance against ‘unseen’ data is the September 2014 measurement
values (hence the September 2014 data set would need to be excluded from the Training/Validation sets as it represents the testing dataset).
Assessment of best performing neural network prediction will be primarily based on the accuracy of the prediction compared to the real world data. Hence the key performance metric that will be used is the Mean Square Error (MSE) for all cases.
20th April 2015: Brief description of the algorithm used and generalisation results.
20th May 2015: Announcement of the competition results
The Completion Winner will be entitled to Free conference registration
Interested researchers are invited to contribute their papers to conference and participate in the Competition. Papers should be submitted electronically through the conference website at http://www.ijcnn.org/
Hissam Tawfik, Liverpool Hope University, UK. Email: email@example.com
David Reid, Liverpool Hope University, UK. Email: firstname.lastname@example.org