| Popis: |
This chapter presents an inductive learning procedure that combines several techniques to generate a fully data‐driven forecasting model. It considers a forecasting method based on appropriate techniques for controlling Artificial neural networks (ANNs) complexity with simultaneous selection of explanatory input variables via a combination of filter and wrapper techniques. Input selection is performed, without user intervention, by applying Chaos theory and Bayesian inference. The challenging problem of rainfall forecasting is employed for showing the robustness of the proposed technique in dealing with different time‐series dynamics. In nonlinear chaotic time‐series analysis, local models are developed via the application of an automatic clustering algorithm based on the rival penalized expectation‐maximization (RPEM) algorithm. Neural network models are estimated, without cross‐validation, relying on data partitioning and Bayesian regularization for complexity control. The proposed forecasting model has been successfully tested using rainfall data from six major hydrographic basins in Brazil. |
