An adaptive hybrid model for monthly streamflow forecasting

This paper suggests a new algorithm for generating Takagi-Sugeno fuzzy systems applied for time series prediction. The model proposed comprises two phases. First, the model structure is initialized in a constructive offline fashion, via an Expectation Maximization algorithm (EM). In the second phase the system is modified dynamically, via adding and pruning operators. At this stage, we propose a recursive learning algorithm, which is based on the EM optimization technique. This online algorithm determines automatically the number of rules necessary at each step. In this way, the model structure and parameters are updated during the adaptive training. The adaptive learning process reduces model complexity and defines automatically its structure providing an efficient model. The proposed approach is applied to build a time series model for monthly streamflow forecasting. The performance of the approach is compared with conventional models used to forecast streamflows. Results show similar errors, however, the suggested model presents a simpler and more parsimonious structure.