Adaptive load consumption modelling on the user side. Contributions to load forecasting modelling based on supervised mixture of experts and genetic programming
- Autores: Francisco Javier Giacometto Torres
- Directores de la Tesis: José Luis Romeral Martínez (dir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2017
- Idioma: español
- Tribunal Calificador de la Tesis: Juan José Cardenas Araujo (presid.), Juan Antonio Ortega Redondo (secret.), Hakan Ergun (voc.)
- Programa de doctorado: Programa Oficial de Doctorado en Ingeniería Electrónica
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
This research work proposes three main contributions on the load forecasting field: the enhancement of the forecasting accuracy, the enhancement of the model adaptiveness, and the automatization on the execution of the load forecasting strategies implemented. On behalf the accuracy contribution, learning algorithms have been implemented on the basis of machine learning, computational intelligence, evolvable networks, expert systems, and regression approaches.
The options for increase the forecasting quality, through the minimization of the forecasting error and the exploitation of hidden insights and miscellaneous properties of the training data, are equally explored in the form of feature based specialized base learners inside of a modelling ensemble structure. Preprocessing and the knowledge discovery algorithms are also implemented in order to boost the accuracy trough cleaning of variables, and to enhance the autonomy of the modelling algorithm via non-supervised intelligent algorithms respectively.
The Adaptability feature has been enhanced by the implementation of three components inside of an ensemble learning strategy. The first one corresponds to resampling techniques, it ensures the replication of the global probability distribution on multiple independent training sub-sets and consequently the training of base learners on representatives spaces of occurrences.
The second one corresponds to multi-resolution and cyclical analysis techniques; through the decomposition of endogenous variables on their time-frequency components, major insights are acquired and applied on the definition of the ensemble structure layout. The third one corresponds to Self-organized modelling algorithms, which provides of fully customized base learner¿s.
The Autonomy feature is reached by the combination of automatic procedures in order to minimize the interaction of an expert user on the forecasting procedure. Experimental results obtained, from the application of the load forecasting strategies proposed, have demonstrated the suitability of the techniques and methodologies implemented, especially on the case of the novel ensemble learning strategy.
