Previsão de carga de Smart Grids: nova formulação por redes neurais da Teoria da Ressonância Adaptativa
Carregando...
Arquivos
Data
2020-03-11
Autores
Silva, Marcela Alexandra da
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Esta pesquisa de doutorado tem por objetivo o desenvolvimento de um sistema previsor de carga de sistemas de energia elétrica baseado no uso de redes neurais artificiais. A previsão de carga tem sido um assunto abordado há muito tempo. Neste sentido, por certo, pode-se afirmar que a literatura técnica atingiu um alto grau de evolução nesta área do conhecimento. Contudo, observa-se que há, ainda, um longo caminho para explorá-la tendo vista às necessidades de atendimento a um setor que evolui constantemente, em especial, aos atuais e aos futuros sistemas elétricos chamados smart grids. Para a realização do planejamento e da operação destes sistemas, várias ações devem ser realizadas: a inserção de modernas tecnologias, uso intenso de dispositivos e de equipamentos eletrônicos, processamento de sinais, inteligência computacional, sistemas de comunicação eficientes, análise e monitoramento das grandezas elétricas, entre outras funções especializadas. Estas ações fazem-se necessárias para tornar exequível o conhecimento do estado do sistema, em regime permanente e, principalmente, em razão da existência de distúrbios elétricos que não são controláveis (manobras acidentais de equipamentos elétricos, acidentes, descargas atmosféricas etc.). Para realizar estas ações, é impreterível o desenvolvimento de metodologias em caráter antecipatório. O principal e indispensável agente antecipatório refere-se à estimação (previsão) de carga futura. A partir do conhecimento de carga futura, pode ser desenvolvida uma série de estudos elétricos (apuração das informações sobre a quantidade de geração necessária para atender a demanda, considerando-se um horizonte predefinido, fluxo de potência, análise de estabilidade, entre outros estudos). Considerando-se que, também, observa-se, na atualidade, um grande desenvolvimento de técnicas para síntese e para a análise de sistemas, em termos de eficiência, precisão e rapidez na obtenção de respostas, é plausível imaginar que o tempo correspondente à execução destas rotinas, ainda em caráter antecipatório, pode se aproximar bastante do tempo real. Assim sendo, nesta pesquisa, dedicar-se-ão os esforços visando contribuir em três ações principais: (1) melhor conhecimento e interpretação das informações contidas na base de dados históricos; (2) explorar a máxima capacidade de realização da rede neural e (3) combinar estes dois objetivos com vistas a proporcionar uma previsão considerando-se um tempo mais próximo do tempo real. Neste sentido, deve-se ressaltar que o tempo futuro, com a máxima aproximação do tempo presente, é bastante crucial, pois à medida que se aproxima do tempo real, perde-se cada vez mais a influência das informações sobre os dados históricos. Neste caso, devem-se contemplar, primordialmente, as medidas advindas do sistema de aquisição de dados. Para a realização deste sistema previsor, usar-se-á uma rede neural da família ART (Adaptive Resonance Theory), mais especificamente a arquitetura supervisionada ARTMAP-Fuzzy. Estas redes neurais têm-se mostrado bastante eficientes, rápidas e versáteis à inclusão de inovações em decorrência de serem arquiteturas estáveis e plásticas, as quais são propriedades diferenciais em relação a outras arquiteturas neurais disponíveis na literatura. A plasticidade constitui-se numa propriedade que possibilita a inclusão das medidas fornecidas pelo sistema de aquisição de dados, cuja idealização pode ser implantada por meio da proposição de uma rotina de treinamento continuado. Visando testar a metodologia proposta, apresenta-se uma aplicação tomando-se como exemplo, uma base de dados disponibilizada por uma empresa internacional do setor elétrico.
This research aims to develop a system for load forecasting of electrical power systems based on the use of artificial neural networks. Load forecasting has been a subject studied for a long time. In this sense, of course, it can be said that technical literature has reached a high degree of evolution in this area of knowledge. However, it is observed that there is still a long way to explore it in view of the needs of serving a sector that is constantly evolving, especially to current and future electrical systems called smart grids. To carry out the planning and operation of these systems, several actions must be carried out: the insertion of modern technologies, intense use of electronic devices and equipment, signal processing, computational intelligence, efficient communication systems, analysis and monitoring of electrical quantities, among other specialized functions. These actions are necessary to make the knowledge of the state of the system feasible, in a permanent regime and, mainly, due to the existence of electrical disturbances that are not controllable (accidental maneuvers of electrical equipment, accidents, lightning, etc.). In order to carry out these actions, it is essential to develop methodologies in anticipation. The main and indispensable anticipatory agent refers to the estimation (forecast) of the future load. From the knowledge of the future load, a series of electrical studies can be developed (verification of the information on the amount of generation needed to meet the demand, considering a predefined horizon, power flow, stability analysis, among other studies). Considering that, also, that a great development of techniques for synthesis and systems analysis is observed, in terms of efficiency, precision and speed in obtaining answers, it is plausible to imagine that the time corresponding to the execution of these routines, still in anticipation, can be very close to real time. Therefore, in this research, efforts will be dedicated to contribute to three main actions: (1) better knowledge and interpretation of the information contained in the historical database; (2) explore the maximum realization capacity of the neural network and (3) combine these two objectives in order to provide a forecast considering a time closer to real time. In this sense, it must be emphasized that the future time, with the maximum approximation of the present time, is quite crucial, because as it approaches real time, the influence of information on historical data is increasingly lost. In this case, the measures arising from the data acquisition system must be considered. For the realization of this predictive system, a neural network of the ART family (Adaptive Resonance Theory) will be used, more specifically the supervised architecture Fuzzy-ARTMAP. These neural networks have been shown to be quite efficient, fast and versatile for the inclusion of innovations due to being stable and plastic architectures, which are differential properties in relation to other neural architectures available in the literature. Plasticity is a property that makes it possible to include the measures provided by the data acquisition system, whose idealization can be implemented through the proposal of a continuous training routine. In order to test the proposed methodology, an application is presented taking, for example, a database made available by an international company of the electricity sector.
This research aims to develop a system for load forecasting of electrical power systems based on the use of artificial neural networks. Load forecasting has been a subject studied for a long time. In this sense, of course, it can be said that technical literature has reached a high degree of evolution in this area of knowledge. However, it is observed that there is still a long way to explore it in view of the needs of serving a sector that is constantly evolving, especially to current and future electrical systems called smart grids. To carry out the planning and operation of these systems, several actions must be carried out: the insertion of modern technologies, intense use of electronic devices and equipment, signal processing, computational intelligence, efficient communication systems, analysis and monitoring of electrical quantities, among other specialized functions. These actions are necessary to make the knowledge of the state of the system feasible, in a permanent regime and, mainly, due to the existence of electrical disturbances that are not controllable (accidental maneuvers of electrical equipment, accidents, lightning, etc.). In order to carry out these actions, it is essential to develop methodologies in anticipation. The main and indispensable anticipatory agent refers to the estimation (forecast) of the future load. From the knowledge of the future load, a series of electrical studies can be developed (verification of the information on the amount of generation needed to meet the demand, considering a predefined horizon, power flow, stability analysis, among other studies). Considering that, also, that a great development of techniques for synthesis and systems analysis is observed, in terms of efficiency, precision and speed in obtaining answers, it is plausible to imagine that the time corresponding to the execution of these routines, still in anticipation, can be very close to real time. Therefore, in this research, efforts will be dedicated to contribute to three main actions: (1) better knowledge and interpretation of the information contained in the historical database; (2) explore the maximum realization capacity of the neural network and (3) combine these two objectives in order to provide a forecast considering a time closer to real time. In this sense, it must be emphasized that the future time, with the maximum approximation of the present time, is quite crucial, because as it approaches real time, the influence of information on historical data is increasingly lost. In this case, the measures arising from the data acquisition system must be considered. For the realization of this predictive system, a neural network of the ART family (Adaptive Resonance Theory) will be used, more specifically the supervised architecture Fuzzy-ARTMAP. These neural networks have been shown to be quite efficient, fast and versatile for the inclusion of innovations due to being stable and plastic architectures, which are differential properties in relation to other neural architectures available in the literature. Plasticity is a property that makes it possible to include the measures provided by the data acquisition system, whose idealization can be implemented through the proposal of a continuous training routine. In order to test the proposed methodology, an application is presented taking, for example, a database made available by an international company of the electricity sector.
Descrição
Palavras-chave
Sistemas elétricos de potência, Previsão de carga global, Redes Neurais Artificiais, Teoria da Ressonância Adaptativa, Treinamento continuado, Electric power systems, Global load forecasting, Artificial neural networks, Adaptive Resonance Theory, Continuous learning