Unleashing the power of artificial neural networks: accurate estimation of monthly averaged daily wind power at Adama wind farm I, Ethiopia

Wind power plays a vital role in the electricity generation of many countries, including Ethiopia. It serves as a valuable complement to hydropower during the dry season, and its affordability is crucial for the growth of industrial centers. However, accurately estimating wind energy poses significant challenges due to its random nature, severe variability, and dependence on wind speed. Numerous techniques have been employed to tackle this problem, and recent research has shown that Artificial Neural Network (ANN) models excel in prediction accuracy. This study aims to assess the effectiveness of different ANN network types in estimating the monthly average daily wind power at Adama Wind Farm I. The collected data was divided into three sets: training (70%), testing (15%), and validation (15%). Four network types, namely Feedforward Backpropagation (FFBP), Cascade Feedforward Backpropagation (CFBP), Error Backpropagation (EBP), and Levenberg–Marquardt (LR), were utilized with seven input parameters for prediction. The performance of these networks was evaluated using Mean Absolute Percentage Error (MAPE) and R-squared (R2). The EBP network type demonstrated exceptional performance in estimating wind power for all wind turbines in Groups GI, GII, and GIII. Additionally, all proposed network types achieved impressive accuracy levels with MAPE ranging from 0.0119 to 0.0489 and R2 values ranging from 0.982 to 0.9989. These results highlight the high predictive accuracy attained at the study site. Consequently, we can conclude that the ANN model’s network types were highly effective in predicting the monthly averaged daily wind power at Adama Wind Farm I. By leveraging the power of ANN models, this research contributes to improving wind energy estimation, thereby enabling more reliable and efficient utilization of wind resources. The findings of this study have practical implications for the wind energy industry and can guide decision-making processes regarding wind power generation and integration into the energy mix.


Introduction
The urgent global challenge of transitioning to sustainable energy systems to mitigate the environmental, economic, and social impacts of fossil fuel consumption has brought renewable energy to the forefront [1][2][3][4][5].Within this global energy landscape, Africa stands as a continent with immense potential to drive the renewable energy revolution [6,7].Among African nations, Ethiopia emerges as a country abundant in low-carbon energy sources such as hydropower, wind, geothermal, and solar [8][9][10][11].However, despite these remarkable opportunities, Ethiopia faces unique obstacles that hinder its full utilization of renewable energy resources [12].Ethiopia heavily relies on hydropower for its energy sector, serving as the primary source of electricity generation.With the second-largest hydroelectric potential in Africa, Ethiopia possesses vast untapped resources.However, currently, only a small fraction of this capacity is being utilized, contributing to approximately 10% of the country's electricity needs.Nonetheless, ongoing hydropower construction projects, with a total capacity of over 6000 MW, hold the promise of significantly enhancing Ethiopia's energy generation capabilities [13,14].
While hydropower offers significant benefits, including renewable and clean electricity generation, its development has not been without challenges.The construction of dams, necessary for hydropower expansion, has given rise to disputes and conflicts, primarily concerning environmental impacts and resettlement of communities [13,15,16].Balancing the energy needs of the nation with the preservation of ecosystems and the well-being of local communities becomes a critical consideration for sustainable hydropower development.Furthermore, Ethiopia's existing power infrastructure and centralized grid systems pose difficulties in extending reliable electricity access to various sectors, including small-scale industries, educational institutions, and healthcare facilities, both in urban centers and remote areas [13].Electric power expansion is challenging with centralized power systems because of distance, dependability, expense, and complexity.Overcoming the limitations of centralized power distribution networks and achieving energy inclusivity require innovative solutions and comprehensive planning.
To unlock Ethiopia's renewable energy potential, diversifying the energy mix becomes imperative.To achieve its renewable energy goals, Ethiopia has recognized wind energy as a viable alternative [17][18][19][20].The complementary nature of wind and hydro power makes them an attractive combination for the country's energy portfolio [17,21,22].In line with this vision, Ethiopia has proposed several wind energy projects with different timelines, including short-term, medium-term, and long-term wind farms with capacities of 970 MW, 1750 MW, and 4000 MW, respectively.However, at present, only three wind farms are operational, generating a total capacity of 324 MW: Ashegoda (120 MW), Adama I (51 MW), and Adama II (153 MW) [8].
Investing in wind energy brings numerous benefits to Ethiopia.Firstly, it can help meet the energy demands of both rural and urban areas.The decentralized nature of wind power allows for the generation of electricity in remote regions that are not connected to the main power grid.This can significantly improve access to electricity in underserved communities, empowering them with reliable and clean energy sources.Furthermore, wind energy can reduce Ethiopia's reliance on fossil fuel imports from other countries.By harnessing its abundant wind resources, Ethiopia can decrease its dependence on costly and environmentally damaging imported fuels.This enhances energy security and helps stabilize energy prices, freeing up financial resources that can be directed towards other developmental priorities.Another significant advantage of wind energy is its contribution to environmental sustainability.Investing in wind power can prevent deforestation by reducing the reliance on wood and charcoal as primary sources of energy in rural areas.This helps preserve valuable ecosystems and biodiversity while mitigating the adverse impacts of deforestation on climate change.
Moreover, wind energy has minimal environmental and health impacts compared to traditional energy sources such as fossil fuels.It does not produce harmful air pollutants or greenhouse gas emissions during electricity generation, leading to improved air quality and better public health outcomes.By reducing reliance on polluting energy sources, wind power can contribute to mitigating climate change and promoting a cleaner and healthier environment for all.Additionally, the adoption of wind energy can alleviate the burdens faced by women and children in rural communities.In many households, women and children are responsible for collecting firewood and biomass fuels, which can be physically demanding and time-consuming.The availability of wind energy can relieve them from these tasks, allowing them to pursue education, engage in incomegenerating activities, and improve their overall well-being.
The safe operation of a power grid heavily relies on the stability and predictability of energy sources.However, wind power, characterized by intermittent and volatile wind speeds, introduces challenges to grid operations, including random fluctuations, instability, and anti-peaking characteristics [23,24].These factors can have a detrimental impact on the quality and reliability of electric power, necessitating accurate forecasting of wind power generation to mitigate these negative effects [14].Wind power forecasting plays a crucial role in maintaining power system reliability, reducing costs, and facilitating informed decision-making by government entities and policymakers [25][26][27].
In the quest for accurate wind power prediction, researchers have explored various statistical, data mining, and machine learning techniques [28][29][30][31][32][33][34][35][36][37][38][39].Among these methods, Artificial Neural Networks (ANN) has gained significant popularity in wind power prediction studies.For instance, ANN models were employed to assess the wind energy output of multiple wind farms in Tamil Nadu, India, utilizing three input variables (wind speed, relative humidity, and generation hours) and achieving low root mean square error (RMSE) and overall percentage error [40].In another study conducted in Rajasthan, India, ANN with the backpropagation algorithm yielded stable mean squared error (MSE) after iterations, using average wind speed, average relative humidity, and generation hours as input parameters [41].Similarly, ANN-based models have been developed for wind power generation forecasting in various wind fields in China, showcasing the efficacy of this approach [42].
Comparative studies have consistently demonstrated the superior performance of ANN models in wind power prediction.In a case study conducted in Tasmania, Australia, the ANN model outperformed the Similar Days approach, exhibiting greater accuracy when evaluated based on the daily mean absolute percentage error (MAPE) [34].Another research conducted in Prince Edward Island, Canada, concluded that ANN outperformed other techniques such as Fuzzy Predictor, Adaptive Neural Fuzzy Inference System, and Committee Machines in wind power prediction [43].Short-term, medium-term, and long-term wind power prediction models have been developed using different methodologies.For example, in Malaysia, researchers extrapolated wind speed data and applied adaptive neuro fuzzy inference systems to create monthly and weekly wind power density prediction models, incorporating metaheuristic techniques such as ant colony optimization and particle swarm optimization [44,45].In Iran, an empirical wind power output model was introduced based on wind speed and ambient temperature data, utilizing both statistical and neural network-based approaches for long-term wind speed prediction [39,45].
Despite the extensive research conducted on wind power prediction, there remains a significant research gap in forecasting wind power specifically in the context of Ethiopia.To address this gap, this study focuses on developing tailored forecasting models using artificial neural networks (ANN) for the 'Adama Wind Farm I.' By harnessing the power of ANN, this research aims to enhance the accuracy of wind power predictions, enabling informed decision-making and optimizing power grid operations for sustainable and efficient wind power generation in Ethiopia.

Study site and data collection
The study was conducted on the Adama wind farm I located 3 km outside Adama town in East Showa Zone, Oromia, Ethiopia (around 8°34′ 38.3″N 39°14′ 35.9″E) (figure 1).This particular location has been identified as one of the prime sites for wind farm establishment in Ethiopia due to its favorable wind resources.Adama wind farm I is an onshore wind farm with a nominal capacity of 51 MW, comprising 34 wind turbines of the Goldwind GW77/1500 model, which are divided into three groups: Group I (GI), Group II (GII), and Group III (GIII).The tower height reaches approximately 65 m, while the blade diameter measures around 76.9 m.The project was completed and became operational in December 2012, successfully connecting to the national grid.For the purpose of this study, data was collected from Adama wind farm I and satellite data from NASA from January 1, 2019, to December 31, 2020.This timeframe was chosen to encompass two consecutive years, providing a substantial amount of historical data for training and validating the wind power prediction.The selected meteorological variables as inputs to the artificial neural network (ANN) for predicting monthly averaged daily wind power (kWh) include wind direction (degree), relative humidity (%), air pressure (kPa), clear Global Solar radiation (CGSR) (kWh/m 2 /day), wind speed (m/s), and minimum and maximum temperature (°C).These features were chosen based on their significance in influencing wind power generation and capturing the relevant environmental conditions at the wind farm site.Air pressure and wind direction are essential factors that affect the behavior and intensity of wind patterns.Relative humidity provides information about the moisture content in the atmosphere, which can impact wind conditions.Temperature variables, including the minimum and maximum values, contribute to understanding the thermal characteristics of the environment, which influence wind behavior.Additionally, CGSR represents the amount of solar radiation available, which indirectly affects wind power generation through its influence on atmospheric conditions.As for the chosen time period (2019-2020), it aligns with the availability of reliable and comprehensive data for the selected meteorological parameters and wind power measurements.By using data from this specific period, we ensure the robustness and accuracy of our analysis.Furthermore, this time frame allows us to capture a sufficient range of seasonal and temporal variations in weather patterns, which are crucial for understanding the dynamics of wind power generation.
By incorporating these diverse meteorological parameters into the prediction model, it enables a comprehensive analysis of their collective impact on wind power generation at the Adama wind farm I.This approach enhances the accuracy and reliability of the forecasting model by considering multiple environmental factors that influence wind power production.
The study was carried out following several steps as detailed below.
Step 1: Data splitting -The collected sample data was divided into three sets: the training data set (70%), the validation data set (15%), and the testing data set (15%).This division allows for the evaluation of different network types, including FFBP, CFBP, EBP, and LR, using distinct data subsets.The training data set is used to train the models and adjust their weights and biases.The validation data set is employed to fine-tune the models and optimize their performance by adjusting hyperparameters.Finally, the testing data set serves as an independent evaluation set to assess the models' generalization ability and measure their predictive accuracy on unseen data.This data splitting strategy helps to ensure that the models are not overfitting or underperforming and allows for the fair comparison of different network types.It enables the identification of the most suitable network type for predicting wind power generation in the Adama wind farm I, Ethiopia.
Step 2: Input features selection: In our study, we conducted a careful selection of input variables for the different types of networks within the Artificial Neural Network (ANN) model.These variables were chosen based on their potential impact on the output variable, which is the prediction of daily averaged wind power.Given the significance of input variables in achieving accurate predictions, their selection was a crucial step in our research.Through this selection process, we ensured that only relevant and influential variables were included in the model.By focusing on the most impactful variables, we aimed to enhance the accuracy and reliability of our predictions.The chosen input variables were optimized and incorporated into the ANN model to improve its overall performance (table 1).
Step 3: Model architecture construction -The architecture of the ANN model was designed, specifying the arrangement and interconnection of its layers and nodes.The model's architecture plays a vital role in capturing the complex relationships between the input and output variables.
Step 4: Model evaluation -The predictive ability of the network type of the ANN model was assessed by comparing the model's predicted outputs with the recorded data.Statistical metrics such as MAPE and R 2 were used to measure the accuracy of the predictions.Lower values of MAPE indicate higher accuracy in the model's predictions, while an R 2 value approaching one signifies greater accuracy.The evaluation process helps to validate the reliability and performance of the ANN model.The general flow chart of the study, illustrating these steps, can be seen in figure 2, providing a visual representation of the research methodology.
For model creation, training, validation and testing, the Data Manager Tool (nntool) in MATLAB R2016a was used, with the input layer consisting of seven neurons representing various factors and the output layer consisting of one neuron representing predicted wind power.The number of neurons in the ANN was incrementally increased until satisfactory prediction accuracy and outcomes were achieved, and the hyperbolic tangent sigmoid math function was employed to transform information in the network layers.The ANN network was executed within the MATLAB environment.

Artificial neural network prediction model
The chosen machine learning method for wind power prediction in this study is the Artificial Neural Network (ANN), which is a powerful computational model inspired by biological neural networks [46][47][48][49][50].The ANN is selected due to its suitability for the research problem and dataset, along with its unique capabilities [51][52][53][54][55].Moreover, ANNs have proven to be highly advantageous as a foundational component within various deep learning methods, including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU) [29,[56][57][58][59][60][61][62].These deep learning architectures have been customized and adapted for specific applications such as image analysis and sequence data processing.One notable advantage of using ANN is its ability to provide transparent interpretation and understanding of the model's inner workings.With structured layers and weighted connections, ANN offers insights into the relationships between input parameters and wind power output.This interpretability enhances the researchers' understanding of how the model arrives at its predictions.Additionally, ANN exhibits flexibility in terms of model complexity.Researchers can adjust the number of layers and neurons within the network to strike a balance between complexity and performance, tailored to the specific research problem at hand.This adaptability ensures that the ANN model captures the necessary information without introducing unnecessary complexity that could potentially lead to overfitting, improving the overall accuracy and generalization capability of the model [63].
Furthermore, ANN demonstrates robust performance even with relatively smaller datasets, making it wellsuited for cases where data availability might be limited.It can effectively learn and model non-linear relationships between input parameters, such as wind direction, wind speed, temperature, humidity, air pressure, and CGSR, to accurately predict monthly averaged wind power.While LSTM, and GRU have their strengths in capturing long-term dependencies in sequential data, ANN's ability to provide meaningful insights, adapt to various data sizes, and deliver comparable performance in wind power prediction tasks justifies its preference in this study.By leveraging ANN's capabilities, the research aims to achieve accurate and reliable predictions of wind power generation at the Adama Wind Farm I in Ethiopia.
In this study, four commonly used neural network training algorithms, namely Feedforward Backpropagation (FFBP), Cascade Feedforward Backpropagation (CFBP), Error Backpropagation (EBP), and Levenberg-Marquardt (LR), were employed for training the data.Two-Layer Feed-Forward Neural Network is employed in this study for its widespread use, balanced complexity and efficiency, strong performance in various domains, effective pattern capture with a moderate number of input features, and suitability for accurate wind power prediction based on meteorological parameters.
FFBP is a widely used algorithm in supervised learning.It utilizes the backpropagation technique to adjust the weights of a feedforward neural network.Information flows from the input layer through hidden layers to the output layer, and the network's outputs are compared to the desired outputs.The error is then propagated backward through the network to update the weights.CFBP is an extension of FFBP that addresses overfitting.It gradually adds neurons and connections during training.Initially, a small network is trained using FFBP, and additional neurons and connections are incrementally added, using existing weights as a starting point.This iterative process continues until the desired performance is achieved.EBP is a general term that encompasses Each of these network types has its own characteristics and suitability for different problems and datasets.The choice of algorithm depends on the specific requirements and constraints of the task at hand.
To develop the prediction models, the ANN was constructed based on seven major climatic factors that directly affect power generation in the area.The governing equation for model development is expressed as: Power F wind direction wind speed minimum temperature maximum temperature relative humidity air pressure and CGSR , , , The development of the ANN framework was carried out using MATLAB version R2016b.The model consisted of seven independent variable matrices, representing the climatic factors mentioned in equation (1), and one dependent variable matrix, which is the power output (figure 3).To ensure the nonlinear formulation and architecture of the ANN, a single hidden layer was incorporated into the modeling process.This hidden layer allows the network to capture and represent complex relationships between the input variables and the power output.

Performance evaluation
In this study, we employed two evaluation metrics, namely Mean Absolute Percentage Error (MAPE) and Coefficient of Determination (R 2 ), to assess the performance of the ANN-based predictions.The MAPE measures the average percentage error relative to the recorded values, with lower values indicating better performance of the ANN model.On the other hand, a higher value of R 2 is desirable, as it indicates a stronger correlation between the predicted values and the actual values, with values closer to 1 being more favorable [64][65][66].

R G G
G G Where G i,rec is recorded wind power, G i, pre is predicted wind power, G rec ¯recorded mean wind power and N-number of data [64][65][66].

Results and discussion
The study utilized different types of ANN networks to predict the monthly averaged daily wind power at Adama Wind Farm I in Ethiopia.The performance of the ANN model was evaluated using two metrics: MAPE and R 2 .After conducting tests and analyzing the results using MAPE and R 2 , it was determined that the best structure for predicting monthly averaged daily wind power was achieved with 17 neurons in the hidden layer.This configuration yielded the lowest MAPE values (ranging from 0.0119 to 0.0489) and the highest R 2 values (ranging from 0.9872 to 0.9989).Table 2 provides a comprehensive evaluation of different network types in the ANN model for each turbine group.
In turbine group GI, the EBP model outperforms other network types, showcasing its excellence in wind power prediction.It achieves an impressive R 2 value of 0.9969, indicating a strong relationship between predicted and actual values.The MAPE value of 0.0119 demonstrates its accuracy with a low average percentage error.The EBP model takes the lead in turbine group GI, followed by FFBP, CFBP, and LR models.Moving to turbine group GII, the CFBP model shines as the most effective network type.It achieves an R 2 value of 0.9978, showing a high level of accuracy.Additionally, the model exhibits a minimum MAPE value of 0.0129, emphasizing its precision.The CFBP model surpasses other network types in turbine group GII, followed by EBP, FFBP, and LR models.In turbine group GIII, the LR model delivers exceptional performance.With an outstanding R 2 value of 0.9989, it indicates an excellent fit between predicted and actual wind power values.The model's MAPE value of 0.0166 further demonstrates its accuracy with a low average percentage error.The LR model takes the lead in turbine group GIII, followed by EBP, FFBP, and CFBP models.
Based on the statistical metrics presented in table 3, all network types in the ANN model prove highly effective in estimating the monthly averaged daily wind power in the study area.However, for each turbine group, specific network types demonstrate superior performance.These findings provide valuable guidance in selecting the most suitable network type for accurate wind power predictions, ensuring reliable and precise modeling results.
Figures 4 to 6 provide an in-depth analysis of the performance of different ANN network types (FFBP, CFBP, EBP, and LR) in predicting the monthly averaged daily wind power compared to the actual recorded data.Specifically, figure 4 shows a comparison between the predicted values using FFBP, CFBP, EBP, LR and the actual values for GI turbine types.The graph reveals a strong agreement between the predicted values obtained from various network types and the actual recorded data, as indicated by the overlapping lines.This suggests that the ANN models successfully capture the complex patterns and dynamics of wind power for Group GI turbines.Furthermore, it is noteworthy that the highest averaged daily wind power is consistently observed in the months of November, December, January, and February.Among all the months analyzed for wind turbine GI type, December emerges as the most wind-intensive month.This finding indicates that wind power generation reaches its peak during December, making it the month with the highest wind speeds and power output for this specific turbine type.This insight highlights the significance of considering the seasonality and variations in wind patterns when planning and optimizing wind energy generation systems.Conversely, the lowest wind power levels are observed in September, May, April, and August.These findings align with the expected seasonal variations in wind power generation.Throughout the remaining months, the wind power levels remain at moderate levels, indicating a relatively stable and predictable pattern.This consistency further reinforces the accuracy and reliability of the ANN models in capturing the wind power trends for Group GI turbines.Overall, figure 4 provides compelling evidence of the ANN models' capability to accurately predict the monthly averaged daily wind power for GI turbine types.The close alignment between predicted and actual values emphasizes the effectiveness of the selected network types (FFBP, CFBP, EBP, and LR) in capturing the intricate dynamics of wind power generation.
The CFBP network type of ANN emerges as the superior model for accurately estimating the monthly averaged wind power for GII.Table 2 and figure 5 demonstrate that this network type consistently outperforms the other network types in terms of fitting the data during the study period.However, it is important to note that all network types of ANN models in this study successfully estimated the monthly averaged daily wind power of GII, as evidenced by the statistical metrics (table 2).
Analyzing the monthly trends, it is observed that the months of April, May, August, and September exhibit the lowest averaged wind power.Conversely, the months of February, January, November, and December showcase the highest levels of monthly averaged wind power.This pattern highlights the seasonal variations in wind power generation for GII.
Of particular interest is the month of February, which stands out as a pinnacle of wind intensity for GII.This notable observation underscores the substantial surge in wind power generation during February, making it the month with the most robust wind speeds and impressive power output for this specific turbine category.The findings highlight the significant potential for harnessing wind energy during the winter season and offer valuable insights for optimizing wind farm operations and capitalizing on the abundant wind resources available during February.
The LR network type of ANN demonstrated superior performance in predicting the monthly averaged wind power for turbine type GIII throughout the entire year (figure 6).This network type consistently provided predictions that closely aligned with the actual values across all months when compared to other network types.Notably, in the months of February, July, September, November, and December, the results obtained from the other network types were less accurate in comparison to LR.This indicates that LR outperformed the alternative network types during these specific months, showcasing its reliability and effectiveness in capturing the wind power patterns for turbine type GIII.
Similar to turbine types GI and GII, the months of February, January, November, and December exhibited the highest monthly averaged wind power for GIII.These months experienced significant levels of wind intensity  and correspondingly higher power output.Conversely, the months of April, May, August, and September displayed the lowest levels of averaged wind power for GIII, indicating relatively less wind activity during these periods.
In order to provide a broader understanding and support the findings of this study, a comprehensive comparison was conducted.This involved analyzing similar research studies that were carried out in different locations and utilized various machine learning algorithms.By examining the results and outcomes of these studies, stakeholders can gain a global perspective on wind power prediction.The details of these comparative studies can be found in table 3.

Conclusion
This paper presents a comprehensive analysis of different network types (FFBP, CFBP, EBP, and LR) in the ANN model for wind power prediction at Adama wind farm I.The study period of 2019-2020 was considered, and the accuracy of the proposed ANN model was evaluated using MAPE and R 2 .
The results obtained indicate that the selected network types, FFBP, CFBP, EBP, and LR, exhibit high effectiveness in accurately estimating wind power in the study area.The MAPE values ranged from 0.0119 to 0.0489, and the R 2 values ranged from 0.9872 to 0.9989, highlighting the robustness of the ANN model.The findings of this research demonstrate the applicability of ANN models in predicting wind power, particularly in regions where data collection equipment may be limited.By utilizing the ANN model, accurate wind power estimates can be obtained, aiding in effective energy production planning and resource management.This study emphasizes the potential of ANN models in the renewable energy sector, specifically for wind power estimation.The accuracy and reliability of the ANN model pave the way for further advancements in optimizing wind farm operations and harnessing wind energy efficiently.
Future research in the field of wind power prediction should focus on exploring and integrating advanced deep learning algorithms.While this study examined various network types within the ANN model, there are promising algorithms such as CNNs, RNNs, and LSTM networks that can enhance prediction accuracy.CNNs can extract spatial features, while RNNs and LSTM networks capture sequential patterns and long-term dependencies.Ensemble methods and hybrid models can further improve accuracy by combining multiple models.Including meteorological, topographical, and environmental variables as input features can enhance predictions by considering additional influencing factors.Scalability and computational efficiency should also be considered, with distributed computing and model compression techniques being potential solutions.By investigating these directions, future research can advance wind power prediction models for more accurate and efficient utilization of wind energy resources.

Figure 1 .
Figure 1.Maps of Ethiopia highlighting the study site Adama in Oromia region.

Figure 3 .
Figure 3. Structure of ANN to predict wind power at Adama Wind Farm I.

Figure 4 .Figure 5 .
Figure 4. Comparison of wind power recorded with predicted using network types of ANN for GI group of turbines at Adama Wind Farm I.

Figure 6 .
Figure 6.Comparison of wind power recorded with predicted using Network types of ANN for GIII group of turbines at Adama Wind Farm I.

Table 1 .
Input and output variables.

Table 2 .
Evaluation of network types of ANN model.

Table 3 .
Comparative analysis of wind power prediction studies.