Temporal Disaggregation of Rainfall Data Using Artificial Neural Networks

Seasonal rainfall in many regions is influenced by global climatic parameters such as El Nino Southern Oscillation (ENSO), La Nina. Seasonal rainfall predictions are based on such Global climatic parameters. However, for operational purposes, such as reservoir operation or river basin management in a larger context, rainfall data at a finer time interval are required. In this paper, a temporal disaggregation model based on Artificial Neural Networks is presented for obtaining rainfall for monthly or shorter time periods from the given seasonal rainfall prediction. A feed forward neural network with back-propagation algorithm for learning was used. A sigmoidal function was used for neuron activation. The training error (RMS error) was measured by squaring the difference between the network and training pattern desired output and summing over all outputs and all training patterns. The methodology is applied to obtain monthly rainfall data from Indian monsoon (JuneSeptember season) rainfall data for a sub-division for which the relation between climatic indices and monsoon rainfall was already established. Monthly rainfall data for 124 years (1871-1994) for Orissa state, India was used for model application. Rainfall data for the first 100 years was used for ANN training and remaining 24 years data was used for testing. The disaggregated monthly rainfall data compared well with observed monthly data. In addition they preserved all the basic statistics such as summing to the seasonal value, cross correlation structure among monthly flows.