Sunday 6 October 2013

The trickiness of forecasting rainfall



Believing  in the weather guy  to keep the golf course dry ? – Forecasting rainfall could be tricky ..

Forecasting rainfall, and at a broader level, forecasting the weather is a problem fraught with challenges. Around the world, various initiatives, such as, Deep Thunder by IBM and Named Peril by Climate Corporation, have been undertaken to address this challenge. In India, the Indian Meteorology Department and the Centre for Mathematical Modeling and Computer Simulation have developed several solutions to address this problem.
The Government of India recently announced a country wide competition, among software companies to forecast rainfall among other similar topics by sharing the historical rainfall and temperature time series data across the different sub-divisions of the Indian subcontinent.

The opportunity to bring in Big Data and the related techniques to help forecast rainfall is immense. Rainfall prediction helps in agriculture, irrigation, disaster management during floods/droughts, and so on. Several attempts have been made, primarily using statistical models, to predict rainfall.

Natural systems, like atmosphere system, exhibit complex dynamical behavior and their data is considered to exhibit nonlinear dynamical, even chaotic behavior. To add to this rainfall data is more sensitive to temporal and spatial variations than other climatic variables. The rainfall time series data, as was provided for the said competition was used to train the neural networks capable of emulating similar behaviors, including deterministic chaos.
So what are the challenges in rainfall forecasting and why these are unique? Chaotic dynamics is associated with extreme sensitivity to initial conditions, exponential divergence of proximal trajectories and have very low predictability horizons. Moreover, most of the variables in the atmospheric state-space are not even measurable, giving rise to computationally intractable behaviors. Hence it is difficult to model and predict atmospheric systems using even higher-order multi-parameter statistical models.

Therefore, machine learning algorithms, specifically recurrent neural networks, capable of emulating the nonlinear dynamical, including chaotic, behaviors have become increasingly popular in the domain of rainfall prediction.

Given the advantages of recurrent neural networks in explaining the nonlinear behavior between the inputs and output, the same were explored to forecast the rainfall of 36 meteorological sub-divisions of India. The model used the past years’ of rainfall data only, to forecast the monsoon rainfall of coming year. The application presented would be providing the forecast for a rolling 18 months period.

A number of algorithms, Elman neural network, Jordan neural network, Radial Basis Function neural network, were applied in a competitive way to obtain the best possible prediction accuracy. The performance of the algorithms was tested across the different regions, categorized by aridity, as the underlying dynamics were expected to change, influenced by geographical location, ocean currents, and so on.
I firmly believe that given the strong emergence of the usage of machine learning algorithms, this challenge of forecasting rainfall with increased degree of accuracy would be a reality in the coming years, and we could believe in the weather guy to keep the golf course dry.

Somjit Amrit is the Chief Business Officer, Technosoft Corporation 
He can be reached at somjit.amrit@technosoftcorp.com

No comments:

Post a Comment