A novel soft computing approach based on fir to model and predict energy dynamic systems

• Autores: Sergio Jurado Gómez
• Directores de la Tesis: Angela Nebot Castells (dir. tes.) , Francisco José Mugica Alvares (codir. tes.)
• Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2020
• Idioma: español
• Texto completo no disponible (Saber más ...)
• Resumen

  • We are facing a global climate crisis that is demanding a change in the status quo of how we produce, distribute and consume energy. In the last decades, this is being redefined through Smart Grids(SG), an intelligent electrical network more observable, controllable, automated, fully integrated with energy services and the end-users. Most of the features and proposed SG scenarios are based on reliable, robust and fast energy predictions. For instance, for proper planning activities, such as generation, purchasing, maintenance and investment; for demand side management, like demand response programs; for energy trading, especially at local level, where productions and consumptions are more stochastics and dynamic; better forecasts also increase grid stability and thus supply security. A large variety of Artificial Intelligence(AI) techniques have been applied in the field of Short-term electricity Load Forecasting(SLF) at consumer level in low-voltage system, showing a better performance than classical techniques. Inaccuracy or failure in the SLF process may be translated not just in a non-optimal (low prediction accuracy) solution but also in frustration of end-users, especially in new services and functionalities that empower citizens. In this regard, some limitations have been observed in energy forecasting models based on AI such as robustness, reliability, accuracy and computation in the edge. This research proposes and develops a new version of Fuzzy Inductive Reasoning(FIR), called Flexible FIR, to model and predict the electricity consumption of an entity in the low-voltage grid with high uncertainties, and information missing, as well as the capacity to be deployed either in the cloud or locally in a new version of Smart Meters(SMs) based on Edge Computing(EC). FIR has been proved to be a powerful approach for model identification and system ’s prediction over dynamic and complex processes in different real world domains but not yet in the energy domain. Thus, the main goal of this thesis is to demonstrate that a new version of FIR, more robust, reliable and accurate can be a referent Soft Computing(SC) methodology to model and predict dynamic systems in the energy domain and that it is scalable to an EC integration. The core developments of Flexible FIR have been an algorithm that can cope with missing information in the input values, as well as learn from instances with Missing Values(MVs) in the knowledge-based, without compromising significantly the accuracy of the predictions. Moreover, Flexible FIR comes with new forecasting strategies that can cope better with loss of causality of a variable and dispersion of output classes than classical k nearest neighbours, making the FIR forecasting process more reliable and robust. Furthermore, Flexible FIR addresses another major challenge modelling with SC techniques, which is to select best model parameters. One of the most important parameters in FIR is the number k of nearest neighbours to be used in the forecast process. The challenge to select the optimal k, dynamically, is addressed through an algorithm, called KOS(K nearest neighbour Optimal Selection), which has been developed and tested also with real world data. It computes a membership aggregation function of all the neighbours with respect their belonging to the output classes.While with KOS the optimal parameter k is found online, with other approaches such as genetic algorithms or reinforcement learning is not, which increases the computational time.