A Concurrent CNN-RNN Approach for Multi-step Wind Power Forecasting
Wind power forecasting helps with the planning for the power systems by contributing to having a higher level of certainty in decision-making. Due to the randomness inherent to meteorological events (e.g., wind speeds), making highly accurate long-term predictions can be extremely difficult. One approach to remedy this challenge is to utilize weather information from multiple points across a geographical grid to obtain a holistic view of the wind patterns, along with temporal information from the previous power outputs of the wind farms. My proposed CNN-RNN architecture combines convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to extract spatial and temporal contexts from multi-dimensional input data in order to make day-ahead predictions. Compared to other methods in the relevant literature, my method incorporates an ultra-wide learning view, combining data from multiple numerical weather prediction models, wind farms, and geographical locations. Specifically, I use a memory-efficient approach where CNN is utilized only for spatial feature extraction, while temporal attributes are diverted towards RNN. This allows me to maximize our spatial layout. Additionally, I experiment with global forecasting approaches to understand the impact of training the same model over the datasets obtained from multiple different wind farms, and I employ a method where spatial information extracted from convolutional layers is passed to a tree ensemble (e.g., Light Gradient Boosting Machine (LGBM)) instead of fully connected layers. Lastly, I consider multitask learning techniques to train a machine learning model to forecast wind power output and classify ramp events simultaneously. The results show that my proposed CNN-RNN architecture outperforms other machine learning models such as LGBM, Extra Tree regressor and linear regression when trained globally, but fails to replicate such performance when trained individually on each farm. I note that the small training size and the lack of high-quality data for such a complex neural network could be a major contributing factor to the lower performance of the CNN-RNN model trained over individual wind farm data. I also observe that passing the spatial information from CNN to LGBM improves its performance, providing further evidence of CNN’s spatial feature extraction capabilities. My results for multitask learning show a slight drop in average forecasting performance, and a slight improvement in average classification performance, however, these changes are not found to be statistically significant.
History
Language
engDegree
- Master of Applied Science
Program
- Mechanical and Industrial Engineering
Granting Institution
Ryerson UniversityLAC Thesis Type
- Thesis