An adaptive ensemble method for quantitative rainfall forecast

In this paper, we have presented an adaptive ensemble method for rainfall forecast. The ensemble is adaptive in sense that the members of the ensemble are trained repeatedly. For this purpose, we have employed strategies in repeated one-step ahead prediction rainfall data. On the other hand, we use diverse models and adapt the weights with which each of these models contribute to the ensemble. We have used, a) multi-layered perceptron network (MLPN), b) Elman recurrent neural network (ERNN), c) radial basis function network (RBFN), and d) generalized neural network (GRNN) as the base models in the ensemble. Each of the base models are trained using soft splitting of the data. The proposed ensemble method has advantages over basic ensemble methods for rainfall forecasting in the sense that the output of this ensemble is an adaptively weighted linear combination of the outputs of the individual models. Moreover, during the test phase the base models are first ranked using least angle regression (LARS). The LARS ranks the variables (i.e., models) according to their predictive performance (i.e., forecasting). In this way, only the higher ranked models are kept reducing the computational complexity of the ensemble. We have set up the case study for the proposed ensemble method on the rainfall series of west central India. The empirical results suggest that the integration of ranking and adaptive fitting of the base models is advantageous than linearly combined ensemble methods in two ways. First, the adaptive ensemble model achieves a competent forecast performance while keeping adaptive property. Second, it has low computational cost as the inefficient base models are discarded while ranking the base models.