LSTM Short-Term Wind Power Prediction Method Based on Data Preprocessing and Variational Modal Decomposition for Soft...

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Signal Processing for Power Systems

Background:

• Accurate real-time power prediction in wind energy is crucial for grid stability and dispatch.
• Traditional methods struggle with the instantaneous measurement challenges of wind power generation.
• Short-term wind power forecasting provides essential data for intraday power grid management.

Purpose of the Study:

• To develop an advanced soft sensor model for improved wind power prediction accuracy.
• To integrate data preprocessing techniques with a Long Short-Term Memory (LSTM) network for enhanced forecasting.
• To leverage Variational Modal Decomposition (VMD) for noise reduction and signal decomposition in wind power data.

Main Methods:

• Data preprocessing including anomaly detection (isolation forest) and multiple imputation for missing values.
• Variational Modal Decomposition (VMD) for decomposing wind power data and reducing noise.
• Long Short-Term Memory (LSTM) network utilizing the Adam optimizing algorithm for predicting individual modal components and reconstructing the final prediction.

Main Results:

• The LSTM network with the Adam optimizer demonstrated superior convergence accuracy.
• Variational Modal Decomposition (VMD) effectively reduced noise and prevented modal aliasing, showing excellent decomposition results.
• The combined LSTM-VMD approach achieved a significant reduction in Mean Absolute Percentage Error (MAPE) by 9.3508%.

Conclusions:

• The proposed soft sensor model integrating LSTM and VMD offers a substantial improvement in wind power prediction accuracy.
• VMD's noise reduction and decomposition capabilities are highly effective when combined with LSTM for time-series forecasting.
• This methodology provides a more reliable reference for intraday power grid dispatch, enhancing the integration of wind energy.