Enhancing wind energy forecasting efficiency through Dense and Dropout Networks (DDN): Leveraging grid search optimization

Abstract

This research investigates wind energy forecasting using a Deep Dense Network (DDN) method enhanced by Grid Search Optimization. It begins with an introduction to wind energy mechanics and machine learning principles, setting the stage for the study's main problem. Further, a comprehensive literature review explores recent machine learning trends for Renewable Energy Sources (RES) output estimation, including deep Neural Networks (NN), Support Vector Regression (SVR), Support Vector Machines (SVM), and other forecasting models, discussing their advantages and disadvantages. The proposed methodology focuses on the Deep Dense Network (DDN) model, detailing its algorithm. The dataset incorporates several variables, such as wind speed, wind direction, temperature, and air pressure and was scaled. A DDN model with eight dense layers of 512, 256, 128, 64, 32, 16, and 8 neurons, each followed by dropout layers (rate 0.4) and using ReLU activation, was designed. The final output layer, with a single neuron, predicts system power. The model was compiled with the Adam optimizer (learning rate 0.1), minimizing MSE and MAE. Early stopping (patience 50 epochs) was employed to prevent overfitting. Grid Search Optimization was applied to fine-tune parameters such as learning rate, dropout rate, batch size, and epochs, improving prediction results The evaluation employs two key metrics: Mean Squared Error (MSE) and Mean Absolute Error (MAE). Together, these metrics provide a comprehensive evaluation of the DDN model's performance by capturing both the average error magnitude (MAE) and emphasizing larger errors (MSE), offering a balanced assessment of prediction accuracy and error distribution. The results demonstrate the model’s capability to converge, indicating effective learning from the data. The application of MSE and MAE metrics substantiates the model's accuracy, with significant reductions in these values reinforcing the proposed approach's validity. Specifically, the MSE decreased from 0.0785 before Grid Search Optimization to 0.0047 after optimization, achieving a 94.013% improvement. Similarly, the MAE reduced from 0.2376 to 0.0548, reflecting a 76.8474% improvement. These substantial enhancements validate the effectiveness of the proposed model. Given the relatively nascent state of renewable energy and deep learning fields, this study offers valuable insights and proposes several directions for future research, establishing a solid foundation for further advancements in this area.