TY - GEN
T1 - Seasonal self-evolving neural networks based short-term wind farm generation forecast
AU - Liu, Yunchuan
AU - Ghasemkhani, Amir
AU - Yang, Lei
AU - Zhao, Jun
AU - Zhang, Junshan
AU - Vittal, Vijay
N1 - Publisher Copyright:
© 2020 IEEE.
PY - 2020/11/11
Y1 - 2020/11/11
N2 - This paper studies short-term wind farm generation forecast. From the real wind generation data, we observe that wind farm generation exhibits the non-stationarity and the seasonality and that the dynamics of non-ramp, ramp-up, and ramp-down events are different across different classes of wind turbines. To deal with such heterogeneous dynamics of wind farm generation, we propose seasonal self-evolving neural networks based short-term wind farm generation forecast. The proposed approach first classifies the historical data into ramp-up and ramp-down datasets and non-ramp datasets for different seasons, and then trains different neural networks for each dataset to capture different wind farm power dynamics. To account for the non-stationarity as well as reduce the burden of hyperparameter tuning, we leverage NeuroEvolution of Augmenting Topologies (NEAT) to train neural networks, which evolves the neural networks using a genetic algorithm to find the best weighting parameters and network topology. Based on the proposed seasonal self-evolving neural networks, we develop algorithms for both point forecasts and distributional forecasts. Experimental results, using the real wind generation data, demonstrate the significantly improved accuracy of the proposed forecast approach, compared with other forecast approaches.
AB - This paper studies short-term wind farm generation forecast. From the real wind generation data, we observe that wind farm generation exhibits the non-stationarity and the seasonality and that the dynamics of non-ramp, ramp-up, and ramp-down events are different across different classes of wind turbines. To deal with such heterogeneous dynamics of wind farm generation, we propose seasonal self-evolving neural networks based short-term wind farm generation forecast. The proposed approach first classifies the historical data into ramp-up and ramp-down datasets and non-ramp datasets for different seasons, and then trains different neural networks for each dataset to capture different wind farm power dynamics. To account for the non-stationarity as well as reduce the burden of hyperparameter tuning, we leverage NeuroEvolution of Augmenting Topologies (NEAT) to train neural networks, which evolves the neural networks using a genetic algorithm to find the best weighting parameters and network topology. Based on the proposed seasonal self-evolving neural networks, we develop algorithms for both point forecasts and distributional forecasts. Experimental results, using the real wind generation data, demonstrate the significantly improved accuracy of the proposed forecast approach, compared with other forecast approaches.
KW - Distributional forecast
KW - Point forecast
KW - Self-evolving neural networks
KW - Short-term wind power forecast
UR - http://www.scopus.com/inward/record.url?scp=85099434815&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85099434815&partnerID=8YFLogxK
U2 - 10.1109/SmartGridComm47815.2020.9303002
DO - 10.1109/SmartGridComm47815.2020.9303002
M3 - Conference contribution
AN - SCOPUS:85099434815
T3 - 2020 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids, SmartGridComm 2020
BT - 2020 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids, SmartGridComm 2020
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2020 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids, SmartGridComm 2020
Y2 - 11 November 2020 through 13 November 2020
ER -