TL;DR
This paper introduces a novel two-stage hybrid model for day-ahead wind power forecasting, combining Hilbert-Huang Transform (HHT), Genetic Algorithms (GA), and Artificial Neural Networks (ANN). This approach significantly improves forecasting accuracy by effectively integrating meteorological data and SCADA records, outperforming existing methods and demonstrating the importance of comprehensive input data.
What Problem Does This Paper Address?
The increasing integration of wind power into electricity grids presents a significant challenge due to the inherent intermittency and volatility of wind speed. This uncertainty complicates grid management for Transmission System Operators (TSOs) and Distribution System Operators (DSOs), impacting grid stability and efficiency. Accurate day-ahead wind power forecasting is crucial for reliable grid operation, enabling better planning for power dispatch, resource allocation, and market participation. Existing forecasting methods often struggle with data gaps or fail to fully leverage both physical (meteorological) and statistical (SCADA) data sources.
Methods
The core of this research is a novel two-stage hybrid forecasting model:
- Stage 1: Wind Speed Forecasting: This stage uses Numerical Weather Prediction (NWP) meteorological data (wind speed, direction, air pressure, temperature, humidity) as input. The data is first decomposed into subseries using the Hilbert-Huang Transform (HHT) to capture complex non-linear patterns. An Artificial Neural Network (ANN), specifically a Multi-Layer Feedforward (MLFF) network, is then trained to predict wind speed at the turbine hub height. To optimize the ANN's performance, Genetic Algorithms (GA) are employed to tune the network's weights and biases, aiming to minimize the mean squared error.
- Stage 2: Wind Power Forecasting: This stage takes the forecasted wind speed from Stage 1 and uses it, along with historical SCADA wind speed and power data, to predict the actual wind power output. Similar to Stage 1, HHT is used for data decomposition, GA for ANN optimization, and ANN for mapping wind speed to power. The final forecasted wind power is obtained by reconstructing the HHT-decomposed subseries.
The paper also investigates the impact of input data selection on forecasting accuracy by testing various combinations of meteorological variables.
Datasets
The study utilized a one-year dataset (May 2014 to April 2015) from a microgrid wind farm in Beijing, China. This included:
- SCADA Data: Ten-minute interval records of actual wind speed and power output.
- NWP Data: Fifteen-minute interval meteorological forecasts from the WRF model, including wind speed, wind direction, air pressure, air temperature, and humidity.
The dataset was split into 85% for training and 15% for validation. The paper also compares its approach against models using different input data subsets to analyze data dependency.
Key Insights and Ideas
- HHT for Signal Decomposition: The Hilbert-Huang Transform proves highly effective in decomposing complex, non-stationary time series data (like wind speed and power) into simpler, more manageable intrinsic mode functions (IMFs). This decomposition allows ANNs to learn more effectively from the underlying patterns.
- GA for ANN Optimization: Using Genetic Algorithms to optimize ANN weights and biases is a robust approach that helps avoid local minima often encountered with traditional backpropagation methods, leading to better generalization and accuracy.
- Two-Stage Hybridization: The two-stage approach, first predicting wind speed and then power, is more effective than a single-stage model. This separation allows for specialized optimization at each step.
- Importance of Comprehensive Input Data: The study clearly demonstrates that including a wider range of meteorological variables (wind speed, direction, air pressure, temperature, and humidity) significantly improves forecasting accuracy compared to using fewer variables. This highlights the value of integrating physical and statistical data.
- Robustness to Data Gaps: While not explicitly tested for missing data in this study, the paper suggests that hybrid approaches incorporating NWP data can offer a practical solution for wind farms with SCADA data gaps.
Key Results
The proposed HHT-GA-ANN model achieved superior forecasting performance compared to benchmark (Smart Persistence) and other AI-based approaches (BP-ANN, GA-ANN, Wavelet-GA-ANN).
- MAPE: The average daily Mean Absolute Percentage Error (MAPE) was 5.54%, representing a significant improvement over other methods. For instance, it was 65.65% better than Smart Persistence.
- NMAE: The average Normalized Mean Absolute Error (NMAE) was 0.48% of the installed capacity.
- Forecast Skill (FS): The model achieved a high average forecast skill of 42.67%, indicating its ability to outperform a simple persistence forecast.
- Input Data Impact: Using all five meteorological variables (wind speed, direction, air temperature, air pressure, and humidity) resulted in the best MAPE (5.54%), showing a 5.78% improvement over using only wind speed.
- Computational Time: The model's average computation time was approximately 30 seconds on a standard personal computer.
Limitations and Future Work
- Limitations:
- The paper focuses on day-ahead forecasting; its performance for longer-term forecasts is not discussed.
- While the approach is suggested as a solution for SCADA data gaps, the paper does not explicitly demonstrate this capability with missing data scenarios.
- The study uses data from a single wind turbine unit; its scalability to larger wind farms with multiple turbines might require further investigation.
- Future Work:
- The authors suggest that the model can be periodically retrained with new input data, implying a need for continuous learning and adaptation.
- Further research could explore the application of this model to larger wind farms and investigate its performance under various weather conditions and data availability scenarios.
Summary updated