Abstract: To address the challenges
of feature selection and the instability of single models in wind power
forecasting, a novel ultra-short-term wind power forecasting method that
integrates multiple LSTM networks with the Q-learning algorithm (Q_L-L-C-A) is
proposed. This method employs the Maximum Information Coefficient (MIC) to
filter the features of wind power data and uses Variational Mode Decomposition
(VMD) to decompose the wind farm power data into multiple frequency modes as
additional features. The filtered and decomposed data are then used as inputs
for the LSTM, CNN-LSTM, and Attention-LSTM network models to make predictions.
Based on this, the Q-learning algorithm is applied to dynamically allocate
weights to the prediction results of the three models to achieve a more
optimized combined forecasting outcome. To validate the forecasting performance
of the proposed Q_L-L-C-A model, actual measurement data from a wind farm is
used as the model input, and comparative experiments are conducted with six
other models. The experimental results indicate that the Root Mean Square Error
(RMSE) and Mean Absolute Percentage Error (MAPE) of the Q_L-L-C-A model are
superior to those of the LSTM, CNN-LSTM, and Attention-LSTM models,
demonstrating that the Q_L-L-C-A model exhibits higher accuracy and stability
in ultra-short-term wind power forecasting.
