MSTL-NNAR: a new hybrid model of machine learning and time series decomposition for wind speed forecasting

Wind speed forecasting is essential for various domains, such as renewable energy generation, aviation, agriculture, and disaster management. However, wind speed is a complex and stochastic phenomenon that exhibits multiple stochastic patterns. Therefore, forecasting methods that can account for the dynamics and uncertainty of wind speed are required. The literature on wind forecasting classifies the time-scale into different categories, such as short-term, medium-term, and long-term. The data used for these categories is usually daily or sub-daily, such as hourly. This implies that high frequency data poses more challenges for the modeling. Machine learning (ML) methods have shown great potential in forecasting wind speed. However, ML methods may not be able to capture the complex and stochastic patterns of time series data, especially when there are multiple seasonal cycles. Therefore, it is often beneficial to use a preprocessing method like time series decomposition before applying ML methods. This article proposes a novel hybrid model that integrates machine learning and time series decomposition for forecasting wind speed. A multivariate seasonal trend decomposition method based on loess (MSTL) is used to preprocess the wind speed data before applying various forecasting methods. MSTL is a generalization of the popular STL method that can split the time series data into a trend component, multiple seasonal components, and a residual component. The seasonal components are considered as deterministic and forecasted using a seasonal naive method that repeats the most recent cycle. For the other non-seasonal components, the neural network autoregression (NNAR) method is employed. The same approach is followed for two common statistical methods: autoregressive integrated moving average and exponential smoothing state space model (ETS). This enables the comparison of the performance of machine learning and statistical forecasting methods based on time series decomposition. The MSTL-NNAR model is also contrasted with the NNAR model without preprocessing to assess the effectiveness of MSTL. The proposed model is applied to two wind speed datasets with different frequencies: daily and hourly. The model is evaluated for both long-term and short-term forecasts by comparing it with other models using various accuracy and bias measures. The results show that the hybrid model, MSTL-NNAR, outperforms all other models in terms of accuracy and bias for both types of data and both time horizons. The results also show that MSTL preprocessing significantly improves the performance of NNAR by reducing its error and bias. The results demonstrate the effectiveness of the hybrid model for forecasting wind speed data that are complex, nonlinear and high-frequency. The model can provide reliable forecasts for different time horizons and different types of data. The model can be useful for various applications, such as wind power generation, weather prediction, and environmental monitoring.