Aplicação do modelo encoder decoder LSTM para previsão de geração fotovoltaica
Data
2024-12-16
Autores
Arruda, Fernando Vasconde de 
Orientadores
Martins, Fernando Ramos
Tipo
Dissertação de mestrado
Título da Revista
ISSN da Revista
Título de Volume
Resumo
Um dos pilares para o desenvolvimento tecnológico da sociedade moderna é a energia elétrica, cujo consumo vem crescendo com os avanços tecnológicos. Grande parte da geração de eletricidade provém de combustíveis fósseis; contudo, as fontes renováveis têm ganho destaque devido aos pactos globais que visam à redução da emissão de gases de efeito estufa com o objetivo de controlar o aumento da temperatura do planeta. O recurso solar é uma alternativa importante para essa transição energética, mas ainda existem desafios tecnológicos que precisam ser superados para ampliar sua utilização. A intermitência da geração fotovoltaica, associada às condições meteorológicas, é um dos principais desafios. Este trabalho aplicou uma metodologia de previsão de geração fotovoltaica para um horizonte de 1 a 3 horas baseada em redes neurais utilizando uma arquitetura conhecida como Encoder-Decoder Long Short-Term Memory (EDLSTM). Esse modelo é altamente eficiente para problemas conhecidos como seq2seq, onde uma sequência de dados é usada como entrada e outra sequência é gerada como saída. O funcionamento do modelo envolve dois componentes: o encoder, que comprime a sequência de entrada em um vetor de comprimento fixo, e o decoder, que descomprime esse vetor para obter a sequência de previsão. Para o treinamento, foram utilizados dados meteorológicos e a série histórica da potência do gerador fotovoltaico localizado na Universidade de São Paulo. A série histórica abrange o período de 2018 a 2022, sendo que os dados de 2018 a 2021 foram usados para treinamento e validação, enquanto os dados de 2022 foram reservados para testes. Foram considerados quatro grupos de atributos de entrada, sendo dois com variáveis atmosféricas e dois com a série histórica de potência. O algoritmo de Grid Search foi implementado para identificar os melhores hiperparâmetros para a topologia das redes. A avaliação do modelo foi realizada utilizando a Raiz do Erro Quadrático Médio (RMSE), o Erro Médio de Viés (MBE) e o Erro Absoluto Médio (MAE), normalizados tanto pela potência nominal do gerador quanto pela média da geração observada (pRMSE, pMBE e pMAE). Além disso, foi utilizado o Skill Score (SS), que compara o desempenho do modelo estudado com um modelo de persistência baseado no índice de claridade de céu claro. Os modelos que utilizaram os atributos de potência como entrada para o treinamento do modelo obtiveram melhor desempenho, com destaque para os modelos EDLSTM P, que apresentou índices de Skill Score variando entre 14,42% e 47,95%, e o EDLSTM PAZ, com índices variando entre 15,16% e 48,67%. Esses dois modelos apresentaram valores positivos de Skill Score em condições de céu claro, nublado e parcialmente nublado, para os três horizontes de previsão.
One of the pillars of technological development in modern society is electricity, the demand for which is growing with technological advances. Much of the generation of this resource comes from fossil fuels; however, renewable sources have gained prominence due to the impacts of climate change and the global pacts aimed at reducing greenhouse gas emissions to control the rise in the planet's temperature. Solar energy is an essential alternative for the transition to renewable sources, but technological challenges still need to be overcome to expand their use. The intermittency of photovoltaic generation, associated with weather conditions, is one of the main challenges. For this reason, this work applied a methodology for forecasting photovoltaic generation for a horizon of 1 to 3 hours, based on neural networks using an architecture known as Encoder-Decoder Long Short-Term Memory (EDLSTM). This model is highly efficient for problems known as seq2seq, where one data sequence is used as input, and another is generated as output. The model's operation involves two components: the encoder, which compresses the input sequence into a vector of fixed length, and the decoder, which decompresses this vector to obtain the prediction sequence. Meteorological data and the historical power series of the photovoltaic generator located at the University of São Paulo were used for training. The historical series covers the period from 2018 to 2022. Data from 2018 to 2021 was used for training and validation, while data from 2022 was reserved for testing. A Grid Search algorithm was also implemented to identify the best set for model hyperparameters. Four groups of input attributes were considered, two with atmospheric variables and two with the historical power series. The model was evaluated using the Root Mean Square Error (RMSE), Mean Bias Error (MBE), and Mean Absolute Error (MAE), normalized by both the generator's peak power and the observed average (pRMSE, pMBE, and pMAE). In addition, the Skill Score (SS) was used, which compares performance with a persistence model developed based on the clear sky index. As a result, the models that used power attributes as input for training the model performed best, with EDLSTM P standing out, with a skill index ranging from 14.42% to 47.95%, and EDLSTM PAZ, with a skill index ranging from 15.16% to 48.67%. In addition, these two models showed positive skill values in clear, cloudy, and partly cloudy sky conditions for all three forecast horizons.
One of the pillars of technological development in modern society is electricity, the demand for which is growing with technological advances. Much of the generation of this resource comes from fossil fuels; however, renewable sources have gained prominence due to the impacts of climate change and the global pacts aimed at reducing greenhouse gas emissions to control the rise in the planet's temperature. Solar energy is an essential alternative for the transition to renewable sources, but technological challenges still need to be overcome to expand their use. The intermittency of photovoltaic generation, associated with weather conditions, is one of the main challenges. For this reason, this work applied a methodology for forecasting photovoltaic generation for a horizon of 1 to 3 hours, based on neural networks using an architecture known as Encoder-Decoder Long Short-Term Memory (EDLSTM). This model is highly efficient for problems known as seq2seq, where one data sequence is used as input, and another is generated as output. The model's operation involves two components: the encoder, which compresses the input sequence into a vector of fixed length, and the decoder, which decompresses this vector to obtain the prediction sequence. Meteorological data and the historical power series of the photovoltaic generator located at the University of São Paulo were used for training. The historical series covers the period from 2018 to 2022. Data from 2018 to 2021 was used for training and validation, while data from 2022 was reserved for testing. A Grid Search algorithm was also implemented to identify the best set for model hyperparameters. Four groups of input attributes were considered, two with atmospheric variables and two with the historical power series. The model was evaluated using the Root Mean Square Error (RMSE), Mean Bias Error (MBE), and Mean Absolute Error (MAE), normalized by both the generator's peak power and the observed average (pRMSE, pMBE, and pMAE). In addition, the Skill Score (SS) was used, which compares performance with a persistence model developed based on the clear sky index. As a result, the models that used power attributes as input for training the model performed best, with EDLSTM P standing out, with a skill index ranging from 14.42% to 47.95%, and EDLSTM PAZ, with a skill index ranging from 15.16% to 48.67%. In addition, these two models showed positive skill values in clear, cloudy, and partly cloudy sky conditions for all three forecast horizons.
Descrição
Citação
ARRUDA, Fernando Vasconde de. Aplicação do modelo encoder decoder LSTM para previsão de geração fotovoltaica. 2024. 182 f. Dissertação (Mestrado Interdisciplinar em Ciência e Tecnologia do Mar) - Universidade Federal de São Paulo, Instituto do Mar, Santos, 2024.