Wind Speed Forecasting for Power Generation Using a Self-Assembling Closed-Loop Recurrent Neural Network

[摘要] This thesis presents the self-assembling recurrent neural network, SFA (Sequential Function Approximation), as a time series forecasting method for wind speed prediction. We compare its multi-step prediction performance against a proven recurrent neural network, NARX (Non-linear Auto-Regressive neural network with eXogenous inputs), on several univariate and multivariate time series, including weather measurements from the Bogdanci Wind Park in Macedonia. Artificial neural networks, such as NARX, require a good deal of trial and error in finding the optimal network configuration. Training these types of networks also comes with high fluctuations in closed-loop prediction performance on each training initialization due to parameter randomization. The SFA method sidesteps these drawbacks while providing comparable or better prediction. This is achieved with the SFA algorithm assembling the input-output mapping by itself to achieve a tolerance set by the user.

[发布机构] Rice University

[效力级别] learning