Avaliação de aplicação de redes neurais artificiais em métodos de medição-correlação-predição

• Autor(a) principal: Velumani, Saran Vikash
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.22/13823
• topic: Wind resource assessment; MCP; Artificial Neural Network; Short term correlation; Wind speed prediction; Avaliação de recursos eólicos; Correlação de curto prazo; Previsão de velocidade do vento

In this study a single artificial neural network (ANN) model was developed to predict the short term mean hourly wind speed and wind direction at target sites using short term mean hourly reference wind data. Standard multi-layered, feed-forward, backpropagation neural networks with single hidden layer architecture was designed using neural network toolbox for MATLAB. The hidden layers and output layer of the network consist of tangent sigmoid transfer function (tansig) and linear transfer function (purelin) as an activation function. Five different sites from Japan, Saudi Arabia, Jordan, France and Russia with different terrain complexity, completely different weather conditions, and different correlation coefficient between reference and target sites were tested. Single model was constructed, and two different approaches were experimented. Approach 1 made use of entire concurrent period dataset, the output values from the model was compared against the three methods: regression, matrix and neural network. Second approach was built on certain period of data and tested on unused data. The purpose behind the fabrication of this approach is to try and understand the neural network model. The results of approach 1 was that the neural network model is able to statistically perform better than other methods and equally well in predicting wind direction sectors. The maximum mean absolute percentage error for NN MATLAB model was found to be 62.5% in Japan to 23.7% in France. The model suffers in predicting the lower wind speeds which explains the distortion in wind frequency distribution and resulting in Power density deviation. The maximum deviation was -18.1% in Jordan and -7.9% in France. The sites in Japan, Saudi Arabia, France and Russia were considered for approach 2. The results were interesting, in case of japan the first month was better than the last month result. Overall the performance of the model was better in case of France followed by Russia site. The maximum deviation of Power density was noticed in case of Japan’s last month scenario -26.6% to minimum of about 3.2% in France and -5.2% was observed in case of Russia. In Saudi Arabia site, the only case where the concurrent period extends over a period of one year, the performance of the model was statistically good but suffers from same problem of previous cases. The deviation in power density was spotted around -21.4%.