TL;DR
This paper introduces a novel CNN-Informer model for predicting photovoltaic (PV) power generation. By combining Convolutional Neural Networks (CNNs) for local feature extraction with the Informer model for long-term temporal dependencies, and incorporating a similar day clustering approach, the method achieves improved accuracy and stability in PV power forecasting, outperforming existing deep learning models.
What Problem Does This Paper Address?
Accurate prediction of photovoltaic (PV) power generation is crucial for the stable and efficient operation of power grids. PV systems are inherently intermittent and their output fluctuates significantly due to weather conditions, making them challenging to integrate into the grid. Traditional power generation is controllable, but solar energy is not. This unpredictability poses challenges for real-time scheduling and grid management. While various methods exist for PV power forecasting, including physical, statistical, and artificial intelligence approaches, there's a continuous need for more accurate and robust models, especially for handling the long-term, complex temporal dependencies in PV data. Existing deep learning models like LSTMs and Transformers can struggle with very long sequences or have high computational complexity.
Methods
The proposed method, termed CNN-Informer, integrates several key techniques:
- Correlation Analysis: To identify the most influential meteorological factors on PV power. Global horizontal radiation was found to be the most significant.
- Similar Day Clustering: The dataset is divided into four categories based on daily weather patterns (sunny, cloudy, cloudy and rainy, rainy) using the Partitioning Around Medoids (PAM) algorithm. This allows the model to learn specific patterns for different weather conditions.
- Convolutional Neural Network (CNN) Feature Extraction: A 1D CNN module is employed to extract local feature information and trends from the time-series data (both PV power and selected meteorological variables). This helps in capturing short-term patterns.
- Informer Model for Temporal Dependencies: The features extracted by the CNN are then fed into the Informer model. The Informer, known for its efficiency in handling long sequences through a ProbSparse self-attention mechanism, is used to capture the long-term temporal dependencies and relationships within the data.
- Combined Architecture: The CNN and Informer are combined, with the CNN acting as a feature extractor that feeds into the Informer's encoder-decoder structure.
Datasets
The study utilizes a dataset from the "Desert Knowledge Australia Solar Centre" spanning from March 11, 2010, to March 8, 2020. This dataset includes variables such as photovoltaic power (Active Power - AP), wind speed (WS), temperature (WTC), relative humidity (WRH), global horizontal radiation (GHR), diffuse horizontal radiation (DHR), wind direction (WD), and weather daily rainfall (WDR). Data was collected at 5-minute intervals between 7:00 and 19:00 daily.
Key Insights and Ideas
- Weather-Specific Forecasting is Key: Clustering data into distinct weather types (sunny, cloudy, etc.) significantly improves prediction accuracy by allowing models to specialize in different environmental conditions.
- CNNs for Local, Informer for Global: Combining CNNs for capturing fine-grained local patterns with the Informer for understanding long-range temporal dependencies creates a powerful synergy for time-series forecasting.
- Global Horizontal Radiation Dominance: The strong correlation between global horizontal radiation and PV power highlights its critical role in forecasting and justifies its inclusion as a primary input.
- Informer's Efficiency for Long Sequences: The Informer model's ability to handle long time series efficiently, thanks to its ProbSparse self-attention mechanism, is crucial for PV forecasting where historical data can span long periods.
Key Results
The proposed CNN-Informer model demonstrated superior performance compared to benchmark models like BP, CNN, and CNN-LSTM across various weather conditions.
- Reduced Errors: The CNN-Informer model achieved the lowest Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE) across all tested weather types. For instance, on sunny days, its MAE was 0.821 kW, MAPE was 0.024, and RMSE was 1.589 kW, outperforming other models.
- Improved Accuracy: The paper reports significant error reductions: MAE decreased by up to 38.1%, MAPE by up to 50.0%, and RMSE by up to 7.5% compared to the BP model. Similar improvements were observed against CNN and CNN-LSTM models.
- Stability and Accuracy: The model showed high accuracy and stability in predicting photovoltaic power generation, especially in challenging weather conditions with significant irradiance fluctuations.
Limitations and Future Work
- Limitations:
- The paper notes that while the CNN-Informer model generally performs well, there can be a certain range of deviation fluctuations in cloudy and rainy days due to the relatively strong and unpredictable weather changes.
- The selection of convolution kernel sizes was optimized for the specific dataset and model architecture; larger kernels might lead to loss of detailed feature information, while smaller kernels might not extract sufficient features.
- Future Work:
- The authors suggest that further research could explore more advanced feature extraction techniques and potentially integrate more meteorological variables or other influencing factors to further enhance prediction accuracy.
- Investigating the model's performance on datasets from different geographical locations and under varying PV system configurations could also be a valuable next step.
Summary updated