| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 11.09 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
O setor energético está a atravessar por um período de profundo processo de mudança estrutural devido à transição energética, na qual a digitalização é um dos pilares desta transição. No pilar da digitalização, é reconhecido o importante contributo da inteligência artificial, nomeadamente no campo da aprendizagem máquina. Contudo, a dificuldade em compreender como muitos dos modelos de aprendizagem máquina obtêm os resultados torna-se um grande desafio, principalmente em processos de tomada de decisão. Muitas das previsões realizadas no setor energético, como por exemplo de consumo de eletricidade, consideram conjuntos de dados do tipo série temporal e utilizam modelos baseados em regressão. Contudo, verifica-se pouca aplicação dos métodos explicativos que contemplam estes dois temas. Este trabalho procura contribuir para melhor compreensão das previsões de consumo de eletricidade obtidas por modelos de aprendizagem máquina, com propostas de geração de explicações destas previsões, através de métodos explicativos. Neste trabalho são exploradas as explicações visuais obtidas de dois métodos explicativos, o LIME e o SHAP, escolhidos para gerar explicações das previsões de consumo de eletricidade de dois modelos de aprendizagem máquina baseados em regressão. Estes métodos explicativos foram selecionados com base no estado da arte apresentado neste trabalho. Foi selecionado um conjunto de dados real, do tipo série temporal, com registos de consumos de três dispositivos existentes em cinco zonas de um edifício e que contribuem para o consumo de eletricidade deste: ar condicionado, tomadas e lâmpadas. Pretende-se avaliar como o uso destas explicações visuais possibilitam a compreensão de quais os atributos do conjunto de dados que os modelos de aprendizagem máquina atribuem maior importância no processo aprendizagem da previsão do consumo de eletricidade. Outro aspeto a ser avaliado é o tempo de processamento da geração das explicações. Os resultados mostram que o SHAP é o mais robusto no sentido em que apresenta sempre os mesmos resultados em diferentes interações, ao contrário do LIME. Contudo, o primeiro apresenta maior degradação no tempo de processamento. Ambos identificam os atributos relativos aos dispositivos ar condicionado e tomadas como aqueles que os modelos de previsão consideram os mais importantes para a previsão do consumo, contudo com diferente ordem de importância em cada método explicativo.
The energy sector is going through a period of profound structural change due to the energy transition, in which digitalization is one of the pillars of this transition. In the digitization pillar, the important contribution of artificial intelligence is recognized, namely in the field of machine learning. However, the difficulty in understanding how many of the machine learning models obtain the results becomes a great challenge, especially in decision-making processes. Many of the forecasts made in the energy sector, such as electricity consumption, consider time series data sets and use regression-based models. However, there is little application of explanatory methods that contemplate these two subjects. This work seeks to contribute to a better understanding of electricity consumption forecasts obtained by machine learning models, with proposals for generating explanations of these forecasts, through explanatory methods. In this work we explore the visual explanations obtained from two explanatory methods, LIME and SHAP, chosen to generate explanations of electricity consumption predictions from two regression-based machine learning models. These explanatory methods were selected based on the state of the art presented in this work. A real time series dataset was selected, with consumption records of three devices existing in five areas of a building and that contribute to its electricity consumption: air conditioning, sockets and lamps. It is intended to evaluate how the use of these visual explanations make it possible to understand which feature of the dataset that machine learning models identify as most important in the learning process of forecasting electricity consumption. Another aspect to be evaluated is the processing time of generating the explanations. The results show that SHAP is the most robust in the sense that it always presents the same results in different interactions, unlike LIME. However, the first one presents greater degradation in the processing time. Both identify the features related to air conditioning devices and taken as those that the forecasting models consider the most important for forecasting consumption, however with different order of importance in each explanatory method.
The energy sector is going through a period of profound structural change due to the energy transition, in which digitalization is one of the pillars of this transition. In the digitization pillar, the important contribution of artificial intelligence is recognized, namely in the field of machine learning. However, the difficulty in understanding how many of the machine learning models obtain the results becomes a great challenge, especially in decision-making processes. Many of the forecasts made in the energy sector, such as electricity consumption, consider time series data sets and use regression-based models. However, there is little application of explanatory methods that contemplate these two subjects. This work seeks to contribute to a better understanding of electricity consumption forecasts obtained by machine learning models, with proposals for generating explanations of these forecasts, through explanatory methods. In this work we explore the visual explanations obtained from two explanatory methods, LIME and SHAP, chosen to generate explanations of electricity consumption predictions from two regression-based machine learning models. These explanatory methods were selected based on the state of the art presented in this work. A real time series dataset was selected, with consumption records of three devices existing in five areas of a building and that contribute to its electricity consumption: air conditioning, sockets and lamps. It is intended to evaluate how the use of these visual explanations make it possible to understand which feature of the dataset that machine learning models identify as most important in the learning process of forecasting electricity consumption. Another aspect to be evaluated is the processing time of generating the explanations. The results show that SHAP is the most robust in the sense that it always presents the same results in different interactions, unlike LIME. However, the first one presents greater degradation in the processing time. Both identify the features related to air conditioning devices and taken as those that the forecasting models consider the most important for forecasting consumption, however with different order of importance in each explanatory method.
Description
Keywords
Explainable Artifical Inteligence Machine Learning Sistemas de Energia Elétrica Electric Power Systems