| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 9.24 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Devido à crescente preocupação com as alterações climáticas e à consequente mudança de paradigma, gradualmente são utilizados mais veículos elétricos e híbridos, como é o caso dos FCHEV (Fuel cell Hybrid Electric Vehicle). Como é característico nos veículos híbridos, a existência de duas fontes de energia faz com que seja crucial otimizar a distribuição de potência, sendo esta a chave para melhorar o desempenho do veículo. Assim, definiu-se como medida principal de desempenho o consumo equivalente em Gasoline Gallon Equivalent, que efetua um tradeoff entre o consumo de hidrogénio e o consumo ponderado de energia da bateria, em função do seu estado de carga. Tendo como principais objetivos o aumento do tempo de vida dos componentes e a redução do consumo do veículo, construiu-se uma estratégia de gestão energética em tempo real, baseada em programação dinâmica, com extração de regras de controlo pela response surface methodology e implementação de machine learning para a identificação dos tipos de ciclo de condução. A estratégia foi construída e simulada em Matlab, partindo da modelação do sistema e da implementação da função DPM (Dynamic Programming Matrix), desenvolvida pelo instituto ETH Zurich para efetuar a programação dinâmica. Seguidamente, foi utilizada a função stepwiselm e a app Regression Learner para extrair as regras de controlo e, finalmente, recorreu-se à app Classification Learner para identificar os ciclos de condução. Toda a estratégia foi complementada com o Matlab Coder, para fazer a transição do algoritmo para linguagem C, suportada pela ECU. Os resultados foram analisados no final de cada fase de implementação, validando a metodologia proposta. Assim, na fase de otimização demonstrou-se que é possível melhorar o consumo equivalente, relativamente ao algoritmo implementado no veículo, obtendo-se uma redução média superior a 15%, sem se demonstrarem alterações significativas no consumo de H2. A partir destes resultados, efetuou-se a extração de regras de controlo, utilizando duas estratégias distintas: regressões não lineares e árvores de decisão. No caso da primeira, não foi possível demonstrar que efetivamente o consumo equivalente é menor, apesar da percentagem de redução desse consumo ser em média superior a zero. No caso da segunda, a robustez do modelo de machine learning demonstrou que em média o consumo equivalente é menor do que no algoritmo atualmente presente no autocarro, sendo que a percentagem de redução em média ultrapassa os 10%. Com ambas as estratégias, as alterações no consumo de H2 não se mostraram significativas. Na fase de reconhecimento do ciclo de condução, utilizou-se uma árvore de decisão que foi analisada para diferentes tempos de decisão, demonstrando-se que com 600 e 300 segundos a identificação apresentou os melhores resultados de accuracy, sendo percetível que para 300 segundos será reduzido o espaço em memória na ECU para armazenamento dos parâmetros de condução, mostrando-se também ser mais preciso em cenários mais semelhantes com a realidade. Finalmente os testes de estrada demonstraram melhorias de 15.7% no consumo equivalente, 24.4% no consumo de H2 e 6.8% no rendimento, com a estratégia que utiliza regressões não lineares. No entanto, o algoritmo mais adequado seria o construído com árvores de decisão, que devido à sua complexidade não foi possível de implementar na ECU.
Due to the growing focus on climate change and its influence in the transportation sector, more and more electric and hybrid vehicles are being used, much like FCHEV (Fuel cell Hybrid Electric Vehicle). Having a double energy source, it is crucial to optimize power distribution which is the key element to improving the vehicle’s performance. As such, for the sake of measuring performance, the Gasoline Gallon Equivalent consumption was used as the main unit of measurement because it jointly considers the energy tradeoff between hydrogen consumption and the estimated battery energy consumption based on its charging status. Considering the components’ effective lifetime and low energy consumption as the main objectives, a live energy management strategy was developed based on dynamic programming, control rules from response surface methodology and the implementation of machine learning which identifies the type of driving cycle. This strategy was developed and simulated with Matlab, starting with system modeling and DPM (Dynamic Programming Matrix) function implementation, developed by the ETH Zurich institute to incorporate dynamic programming. Afterwards the stepwiselm function and the Regression Learner app were used to extract control rules and finally, the Classification Learner app was used to identify the driving cycles. This strategy was complemented with Matlab Coder in order to transition the algorithm to C so as to be supported by the bus’ ECU. The results were analyzed at the end of each implementation phase, validating the proposed methodology. Therefore, in the optimization phase, it was proven that it is possible to improve efficiency, compared to the algorithm already implemented in the vehicle, managing an average improvement of 15%, without causing any significant changes to the H2consumption. From these results, the control rules were extracted using two distinct strategies: non-linear regressions and decision trees. For the first one, it wasn’t possible to demonstrate a lower equivalent consumption, even though the efficiency percentage was on average above zero. As for the second one, because of the complexity and resulting effectiveness of the machine learning model, it was showed that on average, the equivalent consumption is lower than the currently implemented algorithm, showing an average cut in equivalent consumption above 10%. With both strategies, the changes in H2 consumption were not significant. In the driving pattern recognition phase, a decision tree was used and analyzed for different decision response times, showing that with 600 and 300 seconds the presented identification was more accurate, knowing that for 300 seconds the used memory space in the ECU for the driving parameters is reduced, while also being more precise in scenarios closer to reality. Ultimately the road tests showed 15.7% of improvements in the equivalent consumption, 24.45% in H2 consumption and 6.8% in efficiency, with the strategy that uses non-linear regressions. However, the most adequate algorithm using the decision tree, due to its complexity, cannot be implemented in the ECU.
Due to the growing focus on climate change and its influence in the transportation sector, more and more electric and hybrid vehicles are being used, much like FCHEV (Fuel cell Hybrid Electric Vehicle). Having a double energy source, it is crucial to optimize power distribution which is the key element to improving the vehicle’s performance. As such, for the sake of measuring performance, the Gasoline Gallon Equivalent consumption was used as the main unit of measurement because it jointly considers the energy tradeoff between hydrogen consumption and the estimated battery energy consumption based on its charging status. Considering the components’ effective lifetime and low energy consumption as the main objectives, a live energy management strategy was developed based on dynamic programming, control rules from response surface methodology and the implementation of machine learning which identifies the type of driving cycle. This strategy was developed and simulated with Matlab, starting with system modeling and DPM (Dynamic Programming Matrix) function implementation, developed by the ETH Zurich institute to incorporate dynamic programming. Afterwards the stepwiselm function and the Regression Learner app were used to extract control rules and finally, the Classification Learner app was used to identify the driving cycles. This strategy was complemented with Matlab Coder in order to transition the algorithm to C so as to be supported by the bus’ ECU. The results were analyzed at the end of each implementation phase, validating the proposed methodology. Therefore, in the optimization phase, it was proven that it is possible to improve efficiency, compared to the algorithm already implemented in the vehicle, managing an average improvement of 15%, without causing any significant changes to the H2consumption. From these results, the control rules were extracted using two distinct strategies: non-linear regressions and decision trees. For the first one, it wasn’t possible to demonstrate a lower equivalent consumption, even though the efficiency percentage was on average above zero. As for the second one, because of the complexity and resulting effectiveness of the machine learning model, it was showed that on average, the equivalent consumption is lower than the currently implemented algorithm, showing an average cut in equivalent consumption above 10%. With both strategies, the changes in H2 consumption were not significant. In the driving pattern recognition phase, a decision tree was used and analyzed for different decision response times, showing that with 600 and 300 seconds the presented identification was more accurate, knowing that for 300 seconds the used memory space in the ECU for the driving parameters is reduced, while also being more precise in scenarios closer to reality. Ultimately the road tests showed 15.7% of improvements in the equivalent consumption, 24.45% in H2 consumption and 6.8% in efficiency, with the strategy that uses non-linear regressions. However, the most adequate algorithm using the decision tree, due to its complexity, cannot be implemented in the ECU.
Description
Keywords
FCHEV Estratégia de Gestão Energética Programação Dinâmica Fuel Cell Consumo Equivalente Energy Management Strategy Dynamic Programming Equivalent Consumption