A Data Distribution-Based Ensemble Generation Applied to Wind Speed Forecasting

• Diogo M. Almeida UFPE
• Paulo S. G. de Mattos Neto UFPE
• Daniel C. Cunha UFPE

Resumo

Variability in wind intensity and direction causes unstable electricity supply to the power system, making integrating this energy into the electrical system a significant challenge for operations and planning practices. Ensembles can be used as an alternative to address the complex patterns over time in wind speed time series. The appropriate size of the training partition for ensemble models depends on dataset characteristics. And, to enhance model training efficiency, it is important to minimize repetitive data. By maintaining a concise training set, it becomes easier to meet the requirements of software and hardware constraints. Mainly because, there is a growing interest in deploying machine learning models on edge devices. For these reasons, a new method called Local Distribution (LocDist) has been introduced in this paper to predict wind speed, utilizing local pattern recognition based on data distribution. In testing with three wind speed time series, LocDist created a compact training subset with less than 20% of the training data. The Diebold-Mariano hypothesis test was employed to assess the significance of the forecast errors of the proposal compared to individual and bagging methods that use the entire training set. The LocDist method with long short-term memory (LSTM) and gated recurrent unit (GRU) won in 100% of the cases. Additionally, the LocDist with extreme learning machines (ELM) and autoregressive integrated moving average (ARIMA) won or tied in 83% and 66% of the cases, respectively.