Short–term wind speed forecasting in smart grids with machine learning techniques: A case study in Kocaeli-Türkiye

The global surge in energy demand has underscored the shortcomings of conventional energy generation methods, which are increasingly unable to meet this growing need. As a result, there has been a pronounced shift towards renewable energy sources, which are favored for their ability to produce energy without adding to environmental pollution. Among these renewable options, wind energy has gained significant attention due to its renewable, sustainable, and environmentally friendly attributes. It stands out as one of the most promising and effective renewable energy sources within contemporary energy systems. However, despite its potential, wind energy faces several challenges that must be addressed to fully integrate it into existing electricity grids. These challenges include the considerable initial costs associated with wind energy projects and the unpredictable and stochastic nature of wind itself, which complicates its integration into the grid. The variability and intermittent nature of wind make it difficult to ensure a consistent and reliable energy supply, posing significant hurdles for grid stability and efficiency. Addressing these issues is crucial for the widespread adoption and effective utilization of wind energy as a reliable component of global energy systems. This thesis addresses the challenge of wind speed prediction, a critical aspect of integrating wind energy into national electricity grids. Accurate forecasting of wind speed is pivotal for managing the inherent variability of wind power generation and ensuring a stable energy supply. The primary aim of this research is to explore and evaluate the efficacy of artificial neural network (ANN) models for short-term wind speed prediction within the context of a national grid system. This study utilizes a comprehensive dataset from the Kandıra wind farm in Kocaeli, Turkey, to develop and test various ANN models. Three distinct models are investigated: the Nonlinear Autoregressive Neural Network (NARNN), the Nonlinear Autoregressive Neural Network with Exogenous Input (NARXNN), and the Gated Recurrent Unit (GRU). These models are specifically chosen for their capabilities in handling temporal data and capturing complex patterns in wind speed variations. The research involves generating short-term forecasts using these models and comparing their performance against other data-driven methods to assess their accuracy and reliability. By evaluating the effectiveness of these models, this study aims to provide insights into the optimal approaches for wind speed forecasting, thereby contributing to the broader goal of enhancing the integration of wind energy into power grids and improving energy management strategies. A core objective of this thesis is to improve the learning phase of the existing NARNN model by employing hyperparameter optimization. This process involves adjusting various parameters of the model, including the number of hidden layer neurons, delays and training functions, using real-world data. The performance of the NARNN model is evaluated using two key metrics: Mean Squared Error (MSE) and the Determination Coefficient (R²). The results of this study demonstrate that hyperparameter optimization significantly enhances the accuracy of wind speed predictions generated by the NARNN model. Moreover, the optimized NARNN model is compared to its external counterpart, the NARXNN model. It is observed that the NARNN model produces results that are comparable to those of the NARXNN model, while also operating more efficiently in terms of computational time. This indicates that the NARNN model is both effective and time-efficient, making it a suitable choice for short-term wind speed forecasting tasks. In addition to comparing the NARNN and NARXNN models, this thesis also includes an analysis of various learning algorithms used in the training of these models. The results of this analysis reveal that the Bayesian Regularization (BR) algorithm consistently outperforms other algorithms in terms of learning efficacy. However, it is also noted that the Levenberg Marquardt (LM) algorithm, while slightly less accurate than the BR algorithm, offers a much faster execution time, producing results that are nearly as accurate as those of the BR algorithm. This finding highlights the trade-off between accuracy and computational speed, which is an important consideration when selecting a learning algorithm for wind speed forecasting tasks. The study also explores the performance of the GRU model, which is a type of recurrent neural network designed to handle time-series data. The GRU model achieved the lowest MSE of 0.00722, indicating its superior accuracy in minimizing prediction errors. However, when considering the Determination Coefficient (R²), the NARNN model outperformed the GRU model, with the NARNN model achieving an R² value of 0.886 compared to the GRU model's R² value of 0.756. This suggests that while the GRU model excels in minimizing errors, the NARNN model may be better suited to capturing the variability inherent in wind speed data, making it a strong candidate for specific forecasting applications within smart grids. Furthermore, the NARXNN model was evaluated and produced an MSE of 0.2253 and an R² value of 0.904. While this indicates that the NARXNN model performs well in terms of the R² metric, it also suggests that the model exhibits higher error rates compared to the GRU model. These findings underscore the importance of selecting the appropriate model based on the specific characteristics of the dataset and the forecasting task at hand. In conclusion, this thesis underscores the crucial importance of accurate wind speed forecasting for the successful integration of wind energy into contemporary energy systems. The ability to reliably predict wind speed is essential for managing the variability associated with wind power and ensuring a stable and efficient energy supply. This research has demonstrated the significant impact of hyperparameter optimization on enhancing the performance of artificial neural network (ANN) models, with particular emphasis on the Nonlinear Autoregressive Neural Network (NARNN) for short-term wind speed predictions. The application of advanced optimization techniques has proven to improve the model's accuracy and reliability, showcasing the potential for more precise forecasting in wind energy applications. The comparative analysis conducted in this study highlights the strengths and limitations of various models, including the NARNN, Nonlinear Autoregressive Neural Network with Exogenous Input (NARXNN), and Gated Recurrent Unit (GRU). This evaluation provides valuable insights into the trade-offs between forecasting accuracy and computational efficiency, offering a clearer understanding of the practical considerations involved in model selection and deployment. The findings indicate that while more complex models may offer higher accuracy, they often require greater computational resources, necessitating a balanced approach to model development. The contributions of this thesis extend beyond the immediate findings by providing a detailed framework for the application and evaluation of ANN models in wind speed forecasting. The insights gained from this research can inform future studies and guide practitioners in selecting appropriate models for specific forecasting needs. Recommendations for future work include exploring additional hyperparameter tuning strategies, incorporating more diverse datasets, and investigating hybrid models that combine the strengths of different forecasting approaches. By addressing these areas, future research can further advance the field of wind speed prediction and contribute to the efficient integration of wind energy into national and global energy systems.The findings of this study contribute to the ongoing efforts to optimize the integration of renewable energy sources into power systems, ultimately supporting the global transition towards sustainable energy.