IMD uses new method to predict monsoon, hopes to be more accurate

New Delhi, Apr 20 (UNI) The India Meteorological Department's long range forecast for the South West monsoon this year is based upon a newly-adopted statistical forecast system, touted to be more accurate than earlier versions.

It had predicted the South West monsoon rainfall to be 95 per cent of the Long Period Average (89 cm), with a model error of plus/minus five per cent, for the country as a whole.

The IMD had been adopting a two-stage forecast strategy for the south-west monsoon rainfall for the past four years. The first long range forecast for the south-west monsoon season (June-September) rainfall was issued in April using the eight parameter models and the forecast update was issued in June based on the ten parameter models.

Last year, the department had forecast the South West monsoon rainfall to be 93 per cent of the Long Period Average whereas it turned out to be 100 per cent, hence the prediction was off by seven per cent.

The department had since been making consistent efforts to improve the long range forecasting system. It had now developed new statistical models for forecast South West monsoon rainfall (June-September) for the country as a whole, which were being introduced this year.

A new five-parameter statistical forecast system requiring data up to March, would replace the existing eight parameter power regression model for the first forecast in April.

A new six-parameter statistical forecasting system requiring data up to May would replace the existing ten parameter power regression model for the forecast update in June.

There were three major changes in the new statistical forecast models from the eight/ten Parameter models. First, a new smaller predictor data set; second, use of a new non-linear statistical technique along with conventional multiple regression technique; and third, application of the concept of ensemble averaging.

A common weakness of all statistical models was that while the correlations were assumed to remain constant in future, they may, and in fact do, change with time and slowly lose their significance. IMD had embarked an exercise to examine the stability of the predictors and update the list of predictors. This exercise had yielded a new set of eight predictors. For the April forecast, models were developed using a set of five predictors. For the updated forecast in June, models were developed using a set of six predictors, which included three predictors used for the April forecast.

The most important aspect of the new forecast system was the introduction of the concept of ensemble forecasts. In this method, instead of relying on a single model, all the models with all possible combination of predictors were considered. With five predictors, 31 different models were possible.

Out of all possible models, the best few models were selected based on the skill in predicting monsoon rainfall during a common period. For the April forecast, six best models were identified.

Ensemble mean was computed as the weighted average of the best six models thus identified. The weights were proportional to the skill of the models.

For developing the models, two different statistical techniques namely, Multiple Regression (MR) and Projection Pursuit Regression (PPR) were considered. While the MR technique was a conventional linear model technique and more commonly used, the PPR technique was a non-linear technique, used for the first time in forecasting Indian monsoon rainfall. The PPR method was known for its superiority in capturing the non-linear relationships between the predictors and rainfall. Verification of the results with the past data showed that that the ensemble method performed better than the individual models.

The new statistical forecast system had also shown better performance compared to the eight and ten parameter models during the recent years, including the drought years of 2002 and 2004. The model errors of the April and June forecast systems, however, remained as plus/minus five and four per cent respectively.

IMD would update the forecast in June as a part of the second stage forecasts. Separate forecasts for the July rainfall over the country as a whole and seasonal (June-September) rainfall over the four geographical regions of India would also be issued.

UNI