In
this paper, we propose a novel hybrid multi-model approach for
rainfall forecasting. In this multi-model system we have incorporated
an efficient input selection technique, a set of distinct predictive
models with carefully selected parameter settings, a variable selection
method to rank (weight) the models before combining their outputs
and a simple weighted average to combine the forecasts of all the
models. The input selection technique is based on auto correlation
and partial autocorrelation function, the predictive models are
stepwise linear regression, partial least square regression, multivariate
adaptive regression spline, radial basis kernel gaussian process
and multi layer perceptron with quasi Newton optimization. The
model ranking technique is based multi response sparse regression,
which rank the variables (here models) according to their predictive
performance (here forecasting). We have utilized this rank to use
it as the emph{wegiht} in the weighted average of the forecast
combination of the models. We have applied this novel multi model
approach in forecasting daily rainfall of rainy season of Fukuoka
city of Japan. We have used several performance metrics to quantify
the predictive quality of the hybrid model. The results suggest
that the novel hybrid multi-model approach can make efficient and
persistent short term rainfall forecast. |
|
|
|
|
input
selection, variable ranking, weighted forecast combination,
daily rainfall,
short term forecast.
|
|
|