Comparison of Decomposition-based Machine Learning Models for Multi-step Time Series Forecasting of Wind Power Generation

Renewable power generation is increasingly contributing with each passing day to fulfill the increasing load demand. Integration of renewables at larger scale especially wind power with the existing power system poses different challenges for the electrical utilities due to stochastic nature of wind. One of the major challenge is to maintain the balance between power generation from different resources and load demand. Renewable energy forecasting has a pivotal role to play in assuring the flexible and reliable operation of power system. Accurate forecasting models for renewable power generation not only help in load scheduling, unit commitment and economic dispatch but also in scenario modelling of near future which is done to find optimal mix of existing generation from conventional sources and newly added renewable sources. Therefore, this paper compares three decomposition-based machine learning algorithms to assess their performance in multi-step univariate time-series forecasting of wind power. The selected models are STL-ARIMA, CEEMD-BiLSTM, and CEEMDAN-BiLSTM. These models are selected after an extensive literature survey on powerful decomposition strategies and most frequently used statistical and deep learning algorithms for time-series forecasting applications. Accuracy of these models is tested for different time horizons varying from short-term (A day-ahead) to long-term (Three years-ahead). Results show that CEEMDAN-BiLSTM is the most suitable model for short-term wind power generation forecasting with RMSE obtained in the range from 0.64% to 3.12%. On the other hand, STL-ARIMA yields better wind power forecasts in long-term with RMSE in the range from 4.6% to 9.8%.