A Short-Term Wind Speed Multistep-Ahead Hybrid Prediction Model based on ICEEMDAN and Prophet-GRU-NN

Songlin Jiang

Abstract:

As one of the clean and renewable energy, wind power is growing rapidly worldwide.
It is of great significance to accurately predict wind speed in order to serve
wind power generation.

Wind speed prediction is a time series regression problem. Due to the
randomness of short-term wind speed fluctuations as well as its complex
influencing factors, which show non-stationary and nonlinear characteristics,
it is challenging to get predicted alone. Based on previous research, this
paper draws inspiration from numerical weather forecasting methods that
combine the other easily predictable meteorological elements to assist
wind speed forecasting. For each feature, the most advanced ICEEMDAN
decomposition technology is selected, combined with the new Prophet time
series prediction framework proposed by Facebook to analyze the feature.
Based on the inspiration obtained from SVR, the RBF Kernel Principal Component
Analysis technology is also used to reduce the complexity of time series data.
Moreover, with the help of the current latest recurrent neural network
structure GRU, the most high-end GELU activation function, Nadam optimizer,
and Huber loss function are also applied to the neural network to form the
ICEEMDAN-Prophet-GRU model. Finally, a Neural Network is used to integrate
the prediction results of each feature, correct the predicted value of wind
speed, and improve the generalization ability of the model.

We use meteorological data from the actual power station to verify the model.
The results show that by using the multi-feature ICEEMDAN-Prophet-GRU-NN
model proposed in this paper, we can significantly improve the accuracy of
short-term wind speed prediction 12 hours in advance, as the MAPE reaches 8\%,
so it has supremacy.