**3. Materials and methods**

*Aerodynamics*

global scale).

surface temperature differences. As the zone of maximum solar heating migrates with the seasons—moving northward during the Northern Hemisphere summer and southward as winter approaches—the wind patterns that make up the general circulation also migrate latitudinally. Airflow (or wind) can be divided into three broad categories: waves, turbulence, and mean wind. Each can exist in the boundary layer, where transport of quantities such as moisture, heat, momentum, and pollutants is dominated in the vertical by turbulence and horizontal by the mean wind [22]. Each

The earth's highly integrated wind system can be thought of as a series of deep rivers of air that encircle the planet. Embedded in the main currents are vortices of various sizes, including hurricanes, tornadoes, and midlatitude cyclones. Like eddies in a stream, these rotating wind systems develop and die out with somewhat predictable regularity. In general, the smallest eddies, such as dust devils, last only a few minutes, whereas larger and more complex systems, such as midlatitude cyclones and hurricanes, may survive for several days [21]. The scales of atmospheric motion shown in **Table 1** illustrate the three major categories of atmospheric circulation: microscale, mesoscale, and macroscale (synoptic scale and

In short, wind is the movement of air from an area of high pressure to an area of low pressure. In fact, wind exists because the sun unevenly heats the surface of the earth. As hot air rises, cooler air moves in to fill the void. As long as the sun shines, the wind will blow. And wind has long served as a power source to humans [23]. Wind spins the blades, which turn a shaft connected to a generator that produces electricity, in other words, wind turbines convert kinetic energy contained in the wind first into mechanical and then into electrical energy [24]. Wind is a clean source of renewable energy that produces no air or water pollution. And since the wind is free, operational costs are nearly zero once a turbine is erected. Mass production and technology advances are making turbines cheaper, and many govern-

ments offer tax incentives to spur wind-energy development [23].

sustainable source of energy that can successfully replace fossil fuels [6].

Mesoscale 1.0–100 km Tornadoes, waterspouts, dust devils, land/sea

**Scale Typical size Phenomenon Life span** Microscale 0–1.0 km Small turbulent eddies, thunderstorms Seconds to

breeze, mountain/valley breeze

100–5000 km Hurricanes, tropical storms Days to weeks

Longwaves in the westerlies, trade wind Weeks to years

minutes

Minutes to hours/ days

Nowadays, wind turbine technology is considered matured, and the costs of wind energy are low [6]. Industry experts predict that if this pace of growth continues, by 2050 one third of the world's electricity needs will be fulfilled by wind power [23]. Though wind power has performed well in recent years, it also creates a strong environmental impact, such as visual impact, climatic impact, and noise. Although these impacts seem minor when compared with nonrenewable energy, its effect on humans should not be overlooked, due to its potential great development in usage. In short, with proper and supportive policies toward wind power and a good understanding of its environmental impact, wind energy can be a clean and

can exist in the presence of any of the others or separately.

**222**

**Table 1.**

Synoptic scale

Global scale 1000–

40,000 km

*Scales of atmospheric motion (adapted from [20, 21]).*

In this chapter, we use computational intelligence by artificial neural networks for the next-step prediction of one climatic variable: wind speed. ANN was trained to perform the forecasting of 1 h ahead, and then, using it, the trained network was applied to recursively infer the forecasting for the next 6 h of the wind speed (nowcasting), following the methodology explained in [13]. The activation functions that define the outputs of the neurons in terms of their activity levels, inserted in this simulation, were the sigmoidal function in the form of the hyperbolic tangent function (characterized as continuous, increasing, differentiable, and nonlinear) for hidden layers and linear function to the output layer.

To train the RNN and validate the technique, anemometer data (average hourly values of wind speed, wind direction, and temperature) for 1 year (August 08, 2014, and August 07, 2015) are collected by one tower with anemometer installed at height of 101.8 m to Colonia Eulacio (Uruguay), and data (average hourly values of wind speed, wind direction, temperature, humidity, and pressure) for 1 month (November 30, 2015, until December 31, 2015) are collected by one tower with anemometers installed at height of 100.0 m to Mucuri (Brazil), using the same criteria as in [13], namely, 70% for training/validating data and 30% for simulation. The reason for choosing these periods is these are the months with the totality of data available for the realization of this study.

The Mucuri city (Bahia, Brazil) is located at an altitude of 7.0 m in relation to the sea level, and it has a territorial area of 1775 km2 , approximately. Mucuri's anemometer tower is located in a coastal plain, at a distance of 340.0 m from the sea, with latitude 18°1′31.52″ S and longitude 39°30′51.69″ W (**Figure 2**).

As for Colonia Eulacio Tower in Uruguay, according to datum WGS84, it is located at 33o 16′ S, 57o 31′ W [25]. The altitude of the installation location (see **Figure 3**) is approximately 100.0 m, and the location is surrounded by fields with plains; thus, it is characterized by noncomplex terrain. The station is owned by the Administración Nacional de Usinas y Transmissiones Eléctricas (UTE), which is a state-owned company in Uruguay that is responsible for the generation, distribution, and commercialization of electrical energy in the country, as cited in [13].

**Figure 2.** *Location of the Mucuri Tower in Bahia, Brazil.*

**Figure 3.** *Location of the Colonia Eulacio Tower in Soriano Department, Uruguay [21].*

The insertion of meteorological parameters as input data contributes to efficient training of the ANN. Seven different ANN configurations are applied for each site and height; then, a quantitative analysis is conducted, and the statistical results (MAE, MSE, and RMSE) are evaluated to select the configuration that best predicts the real data. The proposed ANN configurations to be analyzed are the following ones. For Mucuri (Brazil) the best ANN configuration was Configuration 1 and for Colonia Eulacio (Uruguay) was the Configuration 4.


**225**

*Study of the Wind Speed Forecasting Applying Computational Intelligence*

percentage error (MAPE), mean absolute error (MAE), mean squared error (MSE), and root mean square error (RMSE). Pearson's correlation coefficient ranges from −1.0 to 1.0. Values close to 0.0 are adequate for the MAE, MSE, and RMSE, values close to 0.0% are adequate for the MAPE, and values close to 1.0 are adequate for the R-squared. The software used to program and perform this computational procedure was MATLAB version 7.10.0 (2010) (personal computer, 8 GB RAM), as

Computational intelligence (CI) is the theory, design, application, and development of biologically and linguistically motivated computational paradigms. Over the last few years, there has been an explosion of research on machine learning and deep learning. Nowadays, deep learning has become the core method for artificial intelligence (AI) [26]. AI is one of the newest fields in science and engineering. AI currently encompasses a huge variety of subfields, ranging from the general (learning and perception) to the specific, such as playing chess, proving mathematical theorems, writing poetry, driving a car on a crowded street, diagnosing diseases, and predicting the conditions of the atmosphere for a given location and time. AI is relevant to any intellectual task; it is truly a universal field [27]. In fact, some of the

In the early days of artificial intelligence, the field rapidly tackled and solved problems that are intellectually difficult for human beings but relatively straightforward for computers—problems that can be described by a list of formal, mathematical rules [28]. Many artificial intelligence tasks can be solved by designing the right set of features to extract for that task, then providing these features to a simple machine learning algorithm. The most widely used artificial neuron model is the perceptron proposed in [29, 30]. This model defines a neuron composed of inputs, a summation and an activation function. The value of each input is multiplied by a weight, and the weighted values of the inputs are summed to yield the result of the sum which is used as the input of the activation function. To teach (train) the neuron, the weights are modified so that the output obtained corresponds to the

The multilayer perceptron (MLP) consists of a system of simple interconnected neurons, or nodes, which is a model representing a nonlinear mapping between an input vector and an output vector. The architecture of a MLP is variable but in general will consist of several layers of neurons. The input layer plays no computational role but merely serves to pass the input vector to the network. The terms input and output vectors refer to the inputs and outputs of the MLP and can be represented as single vectors [31]. Moreover, in relation to recurrent neural network (RNN), the definition is that they are powerful sequence-processing models that are equipped with a memory from recurrent feedback connections. One of the current main challenges of RNN is to dynamically adapt to multiple temporal resolutions and scales in order to learn hierarchical representations in time. Since they operate in discrete time steps and update at every time step, it is generally difficult to learn temporal features that have a significantly different resolution than

Predicting the short-term power output of a wind turbine (wind energy converter) is an important task for the efficient management of smart grids. Shortterm forecasting at the minute scale also is known as nowcasting. By definition nowcasting refers to short lead time weather forecasts, the US National Weather Service specifies zero to 3 h, though forecasts up to 6 h may be called nowcasts by

or R-squared), mean absolute

*DOI: http://dx.doi.org/10.5772/intechopen.89758*

the methodology applied in [13].

(r or Pearson's r), coefficient of determination (R2

**3.1 Computational intelligence and nowcasting**

most successful AI systems are based on CI.

desired value [30].

their input frequency [32].

For statistical analysis of wind speed prediction results at the above sites, the following statistical indicators were applied: Pearson's correlation coefficient

*Study of the Wind Speed Forecasting Applying Computational Intelligence DOI: http://dx.doi.org/10.5772/intechopen.89758*

(r or Pearson's r), coefficient of determination (R2 or R-squared), mean absolute percentage error (MAPE), mean absolute error (MAE), mean squared error (MSE), and root mean square error (RMSE). Pearson's correlation coefficient ranges from −1.0 to 1.0. Values close to 0.0 are adequate for the MAE, MSE, and RMSE, values close to 0.0% are adequate for the MAPE, and values close to 1.0 are adequate for the R-squared. The software used to program and perform this computational procedure was MATLAB version 7.10.0 (2010) (personal computer, 8 GB RAM), as the methodology applied in [13].

## **3.1 Computational intelligence and nowcasting**

*Aerodynamics*

**Figure 3.**

The insertion of meteorological parameters as input data contributes to efficient training of the ANN. Seven different ANN configurations are applied for each site and height; then, a quantitative analysis is conducted, and the statistical results (MAE, MSE, and RMSE) are evaluated to select the configuration that best predicts the real data. The proposed ANN configurations to be analyzed are the following ones. For Mucuri (Brazil) the best ANN configuration was Configuration 1 and for

1.Configuration 1: three layers, nine input nodes (site, Brazil) or seven input nodes (site, Uruguay), nine hidden neurons, and one output node

2.Configuration 2: three layers, nine input nodes (site, Brazil) or seven input

3.Configuration 3: three layers, nine input nodes (site, Brazil) or seven input nodes (site, Uruguay), three hidden neurons, and one output node

4.Configuration 4: three layers, nine input nodes (site, Brazil) or seven input

5.Configuration 5: four layers, nine input nodes (site, Brazil) or seven input

6.Configuration 6: four layers, nine input nodes (site, Brazil) or seven input nodes (site, Uruguay), six hidden neurons (first hidden layer) and three

7.Configuration 7: four layers, nine input nodes (site, Brazil) or seven input nodes (site, Uruguay), one hidden neuron (first hidden layer) and one hidden

For statistical analysis of wind speed prediction results at the above sites, the following statistical indicators were applied: Pearson's correlation coefficient

nodes (site, Uruguay), nine hidden neurons (first hidden layer) and six hidden

nodes (site, Uruguay), six hidden neurons, and one output node

nodes (site, Uruguay), one hidden neuron, and one output node

neurons (second hidden layer), and one output node

neuron (second hidden layer), and one output node

hidden neurons (second hidden layer), and one output node

Colonia Eulacio (Uruguay) was the Configuration 4.

*Location of the Colonia Eulacio Tower in Soriano Department, Uruguay [21].*

**224**

Computational intelligence (CI) is the theory, design, application, and development of biologically and linguistically motivated computational paradigms. Over the last few years, there has been an explosion of research on machine learning and deep learning. Nowadays, deep learning has become the core method for artificial intelligence (AI) [26]. AI is one of the newest fields in science and engineering. AI currently encompasses a huge variety of subfields, ranging from the general (learning and perception) to the specific, such as playing chess, proving mathematical theorems, writing poetry, driving a car on a crowded street, diagnosing diseases, and predicting the conditions of the atmosphere for a given location and time. AI is relevant to any intellectual task; it is truly a universal field [27]. In fact, some of the most successful AI systems are based on CI.

In the early days of artificial intelligence, the field rapidly tackled and solved problems that are intellectually difficult for human beings but relatively straightforward for computers—problems that can be described by a list of formal, mathematical rules [28]. Many artificial intelligence tasks can be solved by designing the right set of features to extract for that task, then providing these features to a simple machine learning algorithm. The most widely used artificial neuron model is the perceptron proposed in [29, 30]. This model defines a neuron composed of inputs, a summation and an activation function. The value of each input is multiplied by a weight, and the weighted values of the inputs are summed to yield the result of the sum which is used as the input of the activation function. To teach (train) the neuron, the weights are modified so that the output obtained corresponds to the desired value [30].

The multilayer perceptron (MLP) consists of a system of simple interconnected neurons, or nodes, which is a model representing a nonlinear mapping between an input vector and an output vector. The architecture of a MLP is variable but in general will consist of several layers of neurons. The input layer plays no computational role but merely serves to pass the input vector to the network. The terms input and output vectors refer to the inputs and outputs of the MLP and can be represented as single vectors [31]. Moreover, in relation to recurrent neural network (RNN), the definition is that they are powerful sequence-processing models that are equipped with a memory from recurrent feedback connections. One of the current main challenges of RNN is to dynamically adapt to multiple temporal resolutions and scales in order to learn hierarchical representations in time. Since they operate in discrete time steps and update at every time step, it is generally difficult to learn temporal features that have a significantly different resolution than their input frequency [32].

Predicting the short-term power output of a wind turbine (wind energy converter) is an important task for the efficient management of smart grids. Shortterm forecasting at the minute scale also is known as nowcasting. By definition nowcasting refers to short lead time weather forecasts, the US National Weather Service specifies zero to 3 h, though forecasts up to 6 h may be called nowcasts by

#### *Aerodynamics*

some agencies [33]. Nowcasting is critical when managing operations of the smart grid, such as system integration, ensuring power continuity and managing ramp rates [34]. In this chapter, nowcasting refers to short-term wind speed forecasting 6 h ahead. In [35] they described and evaluated a proposal for nowcasting wind speed for wind farm locations from historical time series, based on the method of regression by support vectors (in short, nowcasting of wind speed using support vector regression: experiments with time series from Gran Canaria).
