Efficient Multi-Class Wind Speed Prediction Through Genetic Algorithm and Multi-Layer Perceptron Integration

Wind energy is a top-rated and environmentally friendly renewable electricity source, particularly in tropical regions and coastal areas where wind speed is critical for wind power plant design. Wind turbines are significantly influenced by the stochastic, fluctuating, and uncertain nature of wind, which affects their performance. Current wind speed predictions often rely solely on meteorological parameters, leading to prediction errors, increased costs, and computational complexity. This study aimed to identify the most influential meteorological parameters in wind speed prediction. Parameter optimization was conducted using the Genetic Algorithm, and multi-class classification was applied using Multi-Layer Perceptron. Feature selection with the Genetic Algorithm employed binary representation to choose optimal features, with fitness calculations applied to multi-class classification using Multi-Layer Perceptron. The results indicated that the four optimal features affecting wind speed are air temperature, air pressure, wind direction, and dew point temperature. Model evaluation using these four optimal features yielded an accuracy of 56%, precision of 51%, recall of 56%, F1-Score of 46%, and ROC-AUC Score of 77%, with a computation time of only 10.6 seconds. The low F1-Score and accuracy, as indicated by sensitivity and specificity evaluations, were attributed to class imbalance, impacting the overall results. Further testing with adjusted data showed an average accuracy of 85%. Despite data imbalance challenges, using selected feature subsets significantly improved accuracy and reduced computation time in predicting wind speed compared to using all original features.