A Machine Learning-Based Sustainable Energy Management of Wind Farms Using Bayesian Recurrent Neural Network

Aisha Blfgeh, Hanadi Alkhudhayr

The sustainable management of energy sources such as wind plays a crucial role in supplying electricity for both residential and industrial purposes. For this, accurate wind data are essential to bring sustainability in energy output estimations for wind stations. The choice of an appropriate distribution function significantly affects the actual wind data, directly influencing the estimated energy output. While the Weibull function is commonly used to describe wind speed at various locations worldwide, the variability of weather information across wind sites varies significantly. Probabilistic forecasting offers comprehensive probability information for renewable generation and load, assisting decision-making in power systems under uncertainty. Traditional probabilistic forecasting techniques based on machine learning (ML) rely on prediction uncertainty derived from previous distributional assumptions. This study utilized a Bayesian Recurrent Neural Network (BNN-RNN), incorporating prior distributions for weight variables in the RNN network layer and extending the Bayesian networks. Initially, a periodic RNN processes data for wind energy prediction, capturing trends and correlation characteristics in time-series data to enable more accurate and reliable energy production forecasts. Subsequently, the wind power meteorological dataset was analyzed using the reciprocal entropy approach to reduce dimensionality and eliminate variables with weak connections, thereby simplifying the structure of the prediction model. The BNN-RNN prediction model integrates inputs from RNN-transformed time-series data, dimensionality-reduced weather information, and time categorization feature data. The Winkler index is lower by 3.4%, 32.6%, and 7.2%, respectively, and the overall index of probability forecasting pinball loss is reduced by 51.2%, 22.3%, and 10.7%, respectively, compared with all three approaches. The implications of this study are significant, as they demonstrate the potential for more accurate wind energy forecasting through Bayesian optimization. These findings contribute to more precise decision-making and bring sustainability to the effective management of energy systems by proposing a Bayesian Recurrent Neural Network (BNN-RNN) to improve wind energy forecasts. The model further enhances future estimates of wind energy generation, considering the stochastic nature of meteorological data. The study is crucial in increasing the understanding and application of machine learning by establishing how Bayesian optimization significantly improves probabilistic forecasting models that would revolutionize sustainable energy management.