Seasonal rainfall hindcasting using ensemble multi-stage genetic programming

Rainfall hindcasting is one of the most challenging tasks in the hydrometeorological forecasting community. The current ad hoc data-driven approaches appear to be insufficient for forecasting rainfall. The task becomes more difficult, when the forecasts are over a long period of time. To increase the accuracy of seasonal rainfall hindcasting, this paper introduces an ensemble evolutionary model that integrates two genetic programming techniques: gene expression programming (GEP) and multi-stage genetic programming (MSGP). To demonstrate the development and validation procedures of the new model, the rainfall data from the Antalya meteorology station was used. The model performance was evaluated in terms of different statistical measures and compared with that of the state-of-the-art gradient boosted decision tree (GBT) model developed as a reference model in this study. The performance results during testing showed that the proposed ensemble model has increased the seasonal forecasting accuracy of the GEP and MSGP models up to 30%. The GBT was found comparable to the proposed model during training period; however, it drastically underestimated extreme wet seasons during testing.

Automated summary

This paper introduces an ensemble evolutionary model to improve the accuracy of seasonal rainfall hindcasting. By combining gene expression programming and multi-stage genetic programming techniques, the model increased the forecasting accuracy by 30% during testing, demonstrating its effectiveness in comparison with the state-of-the-art gradient boosted decision tree model.