Performance of Artificial Neural Network in Nowcasting Summer Monsoon Rainfall: A case Study

Rainfall prediction has always remained challenging for meteorologists and forecasters due to the complex physics involved. Prediction is important as excess or deficient rainfall has adverse effects on the agriculture sector which in turn drives other economic sectors. Advances in technology have made it possible to have a web of weather stations from where data can be collected frequently in the form of numbers, images, graphs etc. With such availability of data, artificial intelligence has always been a choice of researchers for solving this complex problem. Artificial Neural Network (ANN), a data-driven approach is used here to predict the daily summer monsoon rainfall over Shivajinagar region (18.5308N, 73.8475E). Feed Forward Neural Network is employed for the rainfall prediction. Weather parameters are used as inputs. Number of inputs, number of nodes and number of layers are varied and each model is tested for unseen data. It was found that selection of inputs is important in the case of multivariate time series forecasting using ANN. Increasing the number of layers does not always help to increase accuracy. Performance of all the trained networks is tested for the daily summer monsoon rainfall of the year 2008. The predicted rainfall successfully followed an increase and decrease in the observed rainfall with the Mean Absolute Error of 4.6. A new paradigm for comparing the network performance is used here which is maximum and minimum rainfall prediction capability. It is found that a single hidden layer network with all-weather parameters as inputs has the ability to predict rainfall.