Estimation of wind power in coastal areas using a Model based on the learning of a Multilayer Perceptron: Case of Douala, Cameroon

Abstract

In recent years, the scientific community has been concerned about the threat of global warming. This phenomenon is due to the increase of greenhouse gas emissions greenhouse gas emissions due to human activity. Renewable energies present themselves as a potential solution to reduce to reduce greenhouse gas emissions. Among the promising means of production As part of its policy of promoting renewable energy, the Cameroonian government launches each year projects in this area, hence the need to study the various sites and according to the parameters that are most often stochastic, hence the problem of choosing the type of wind power choose and implement. Thus this paper proposes a method of estimating the power produced according to the wind speed data of the scale coefficient and the shape on the station P/30 of Douala of the site of ASECNA - Douala Cameroon over a period of one year which presents different characteristics on the four seasons of the coastal areas of Cameroon by using the distribution of Weibull and by proposing another method using artificial intelligence ; This instability offers the opportunity to study other methods of power estimation using, as in this work, a multilayer perceptron type neural network. Based on the Weibull parameters, the power estimation is done by both approaches according to the different coastal seasons: hard dry season, short rainy season, short dry season and long rainy season. In addition, the form factor and scale factor fluctuated over the year from 1.36 to 1.94 and from 2.74 m/s to 3.80 m/s for different periods respectively. It was found that the average wind speed is 1.309 m/s, the average power for this site is 289.46 MW, and the months of March and July have high powers because the winds are warmer in these periods. For estimation we used a multi-layer perceptron consisting of: 03 input layers (wind speed, form factor and scale factor), 02 hidden layers of 10 neurons each and one output layer (wind turbine power), for training we used the gradient back-propagation algorithm using Matlab software. After an average of 200 training runs and a training step of 0.001, we obtained an RMSE for each of the four seasons of 0, 0065361; 0.00165361; 0.00052543; 0.0000011564. It was concluded that the algorithm improves the accuracy of power estimation by the MLP model and can be recommended for wind turbine power estimation.