Soft computing approaches in ocean wave height predicction for marine energy applications
- Autores: Juan Miguel del Pino González
- Directores de la Tesis: Lucas Cuadra Rodríguez (dir. tes.)
- Lectura: En la Universidad de Alcalá ( España ) en 2018
- Idioma: inglés
- Tribunal Calificador de la Tesis: Antonio José Caamaño Fernández (presid.), Sancho Salcedo Sanz (secret.), Carlos Casanova Mateo (voc.)
- Programa de doctorado: Programa de Doctorado en Tecnologías de la Información y las Comunicaciones por la Universidad de Alcalá
- Materias:
- Enlaces
- Tesis en acceso abierto en: TESEO
- Resumen
- español
El objetivo de esta tesis consiste en investigar el uso de técnicas de Soft Computing (SC) aplicadas a la energía producida por las olas o energía undimotriz. Ésta es, entre todas las energías marinas disponibles, la que exhibe el mayor potencial futuro porque, además de ser eficiente desde el punto de vista técnico, no causa problemas ambientales significativos. Su importancia práctica radica en dos hechos: 1) es aproximadamente 1000 veces más densa que la energía eólica, y 2) hay muchas regiones oceánicas con abundantes recursos de olas que están cerca de zonas pobladas que demandan energía eléctrica. La contrapartida negativa se encuentra en que las olas son más difíciles de caracterizar que las mareas debido a su naturaleza estocástica. Las técnicas SC exhiben resultados similares e incluso superiores a los de otros métodos estadísticos en las estimaciones a corto plazo (hasta 24 h), y tienen la ventaja adicional de requerir un esfuerzo computacional mucho menor que los métodos numérico-físicos. Esta es una de las razones por la que hemos decidido explorar el uso de técnicas de SC en la energía producida por el oleaje. La otra se encuentra en el hecho de que su intermitencia puede afectar a la forma en la que se integra la electricidad que genera con la red eléctrica. Estas dos son las razones que nos han impulsado a explorar la viabilidad de nuevos enfoques de SC en dos líneas de investigación novedosas.
La primera de ellas es un nuevo enfoque que combina un algoritmo genético (GA: Genetic Algorithm) con una Extreme Learning Machine (ELM) aplicado a un problema de reconstrucción de la altura de ola significativa (en un boya donde los datos se han perdido, por ejemplo, por una tormenta) utilizando datos de otras boyas cercanas. Nuestro algoritmo GA-ELM es capaz de seleccionar un conjunto reducido de parámetros del oleaje que maximizan la reconstrucción de la altura de ola significativa en la boya cuyos datos se han perdido utilizando datos de boyas vecinas. El método y los resultados de esta investigación han sido publicados en: Alexandre, E., Cuadra, L., Nieto-Borge, J. C., Candil-García, G., Del Pino, M., & Salcedo-Sanz, S. (2015). A hybrid genetic algorithm¿extreme learning machine approach for accurate significant wave height reconstruction. Ocean Modelling, 92, 115-123.
La segunda contribución combina conceptos de SC, Smart Grids (SG) y redes complejas (CNs: Complex Networks). Está motivada por dos aspectos importantes, mutuamente interrelacionados: 1) la forma en la que los conversores WECs (wave energy converters) se interconectan eléctricamente para formar un parque, y 2) cómo conectar éste con la red eléctrica en la costa. Ambos están relacionados con el carácter aleatorio e intermitente de la energía eléctrica producida por las olas. Para poder integrarla mejor sin afectar a la estabilidad de la red se debería recurrir al concepto Smart Wave Farm (SWF). Al igual que una SG, una SWF utiliza sensores y algoritmos para predecir el olaje y controlar la producción y/o almacenamiento de la electricidad producida y cómo se inyecta ésta en la red. En nuestro enfoque, una SWF y su conexión con la red eléctrica se puede ver como una SG que, a su vez, se puede modelar como una red compleja. Con este planteamiento, que se puede generalizar a cualquier red formada por generadores renovables y nodos que consumen y/o almacenan energía, hemos propuesto un algoritmo evolutivo que optimiza la robustez de dicha SG modelada como una red compleja ante fallos aleatorios o condiciones anormales de funcionamiento. El modelo y los resultados han sido publicados en: Cuadra, L., Pino, M. D., Nieto-Borge, J. C., & Salcedo-Sanz, S. (2017). Optimizing the Structure of Distribution Smart Grids with Renewable Generation against Abnormal Conditions: A Complex Networks Approach with Evolutionary Algorithms. Energies, 10(8), 1097.
- English
Among all the available marine energies, wave energy is the most used because, in addition to being ecient from a technical viewpoint, it does not cause significant environmental problems. Its ecient conversion is based on a variety of wave energy converters (WECs), which transform the kinetic energy of waves into electric energy by means of either the vertical oscillation of waves or the linear motion of waves. Its practical importance lies in two facts: 1) it is about 1000 times denser than wind energy, and 2) there are many ocean regions with strong wave resources that are near populated zones demanding electric energy. The counterpart is that waves are more dicult to characterize than tides because of the stochastic nature of waves (although they are less variable on an hourly basis than wind). Such characterization can be carried out essentially through two families of techniques: physical models and data-driven approaches. Physical models are based on the wave energy balance equation, a di↵erential equation which is solve numerically. They are more accurate in forecasting over large spatial and temporal domains (within a window of a few days) but at the expense of requiring a huge amount of data and computational e↵ort. This is not the case of data-driven models. These are based on data series measured by buoys, radars, or satellites. Data-driven models involve statistical regressive methods (such as auto regressive moving average) and Soft Computing (SC) approaches (neural, fuzzy, and evolutionary methods). SC techniques exhibit similar and even superior results than statistical methods in short term estimations (up to 24 h), and have the additional bonus of requiring much smaller computational cost than numerical-physical methods. Furthermore, the intermittency su↵ered by massive renewable energies (mainly solar, wind and wave energies), in general, and wave energy, in particular, is another problem that must be tackled in the e↵ort of eciently integrating wave energies in the electricity network (or power grid).
These are the reasons that have compelled us to explore the feasibility of novel SC approaches in wave energy applications. Specifically, in this thesis we have focused on two research lines with applications in wave energy and its integration in the power grid.
In the first research line we have centered on characterizing the wave energy resource. In particular, we have tackled a problem of reconstructing an importantparameter used in wave energy called “significant wave height”, Hs. Specifically, we have focused on a kind of problem which consists in reconstructing the significant wave height Hs at the location of an out-of-operation measuring buoy by using wave parameters from other nearby buoys. This situation is quite common in certain occasions in which, due to a storm or other type of accident, a buoy is destroyed or deactivated, so that those wave data that had to be measured by such buoy at that particular location are lost. Data at near buoys can be used to reconstruct the missing data that had to be measured in the damaged buoy. This reconstruction isimportant because, as mentioned, Hs, plays a key role in the design and operation of WECs. We have faced the problem of filling up missing values of Hs within the framework of Machine Learning (ML), in a two-step process, which has led to two contributions. The first one consists in designing a hybrid evolutionary algorithm that selects, among the available wave parameters (from the nearby buoys), a smaller subset FnSP with nSP parameters that minimizes the Hs reconstruction error. For doing this, we have proposed a novel approach in marine energy applications consisting of a Genetic Algorithm (GA) that computes the fitness of the candidate individuals (trial solutions) in each generation by using an Extreme Learning Machine (ELM). In this context, the key advantage of the ELM when compared to other ML approaches (Neural Networks, or Support Vector Machines, for instance) is that ELMs learn very fast, this being essential in population-based evolutionary algorithms such as GAs. This is why we have hybridized the ELM with the GA in the detriment of other alternative ML regressors. The proposed hybrid GA-ELM method generates a subset FnSP of nSP parameters that minimizes the root mean square error of Hs reconstruction, RMSE(Hˆs)(m). In the e↵ort of testing its performance in two di↵erent coastal regions, we have explored two case studies: one in the Caribbean Sea, and the other, in the West Atlantic coast nearby Florida. The results suggest that: The proposed GA-ELM algorithm works very well in the sense that it selects a very reduced subset of parameters (nSP = 10 parameters) among the available 60 parameters. Using 5 6 nSP 6 10 parameters lead to small reconstruction errors: RMSE(Hˆs)Caribbean . 0.50 m, and RMSE(Hˆs)Atlantic . 0.75 m.
The selected wave parameters in subset FnSP assist other ML regressors Extreme Learning Machines, Support Vector Regression (SVR), and Gaussian Process Regression (GPR) in reconstructing Hs. All the ML method explored have Hs reconstruction errors below 1m in the two di↵erent locations studied: RMSE(Hˆs) < 0.5 m in the Caribbean Sea, and RMSE(Hˆs) < 0.75 m in the West Atlantic scenario.
As a general conclusion, the twofold approach explored in our first publication seems to be a feasible tool to fill missing wave parameter values by using data from neighbor buoys.
The second research line explored in this Thesis has been motivated because of the existence of two mutually linked issues. The first one is related to the way in whichthe WECs must be electrically connected to each other (forming a WEC farm) to supply enough electrical energy. The second aspect is related to how to eciently connect the WEC farm with the on-shore power grid. This is because the variability and intermittency of wave energy can a↵ect the stability of the power grid. This second problem can be encompassed within a broader conceptual framework, which is common to all massive renewable energies (REs). In this wider framework, the Smart Grid (SG) paradigm aims at integrating the current growing number of distributed renewable energy-based generators, without significantly a↵ecting the stability of the power grid. The novelty of our approach is twofold, in the sense of optimizing the robustness of a distribution SG (connecting nodes that generate, consume or store electricity) against “abnormal operating conditions” –for instance, the breakdown or the operation stop of a WEC (or a set of them) caused by a storm– by (1) using an Evolutionary Algorithm, (EA) that optimizes the structure of the SG modeled by (2) applying concepts from Complex Networks (CN) Science. Our approach takes advantage of some important properties of the SG paradigm: a smart grid allows for the bidirectional exchange of electric energy at the local scale and aims at supplying reliable and safe electric power by eciently integratingdistributed RE generators using smart sensing and communication technologies. Thanks to the ecient integration of distributed REs in the grid, electricity consumers canalso become producers (“prosumers”), helping end-users obtain economic benefits by selling the energy generated in excess. In a similar abstraction, a Smart Wave Farm (SWF), formed by WECs and energy storing devices, can be viewed as a set of nodes that exchange electric energy at local scale. In this context, we have modeled the SG as an undirected graph so that each link (electric cable) allows for the bidirectional exchange of electric energy between nodes.
Aiming at optimizing the structure of such SG against abnormal conditions, we have proposed a novel objective function (to be minimized) that combines cost elements, related to the number of electric cables, and several metrics that quantify properties that are beneficial for the SG (energy exchange at the local scale and highrobustness and resilience). The optimized SG structure is obtained by applying an EA in which the chromosome that encodes each potential graph (or individual) is the upper triangular matrix of its adjacency matrix. This allows for fully tailoring the crossover and mutation operators to such encoding. Since small-networks have been found to be beneficial for SGs, we have proposed a domain-specific initial population that includes both random networks and small-world networks. This assists the proposed EA to converge quickly.
The experimental work points out that the proposed method performs well and generates an optimum, synthetic, non-sparse-like small-world structure that leads to beneficial properties such as improving both the energy exchange at the local scale and the robustness and resilience. Specifically, the optimum topology fulfills a balance between moderate cost and robustness against abnormal conditions. We would like finally to emphasize two aspects of the second publication of this Thesis: The proposed approach should be considered as a high level analysis and planning tool in the e↵ort of estimating to what extent the smart grid topology can su↵er from vulnerabilities. It cannot and does not intend to replace the conventional methods used by power engineers. In fact, the low level, detailed design must be carried out using electrical engineering techniques.The model is suciently general to be applied to any set of generators and loads (consuming energy) as well as to the Smart Wave Farm system formed by WECs, energy storing devices, and the connection(s) to power grid.
- español
