TL;DR
This paper introduces a novel unsupervised clustering approach (OCCUR) to group daily solar irradiance patterns, followed by a pattern recognition step (SVM-PR) to classify incoming days. This allows for tailored forecasting using a multi-model (M3) framework, significantly improving accuracy over traditional methods.
What Problem Does This Paper Address?
Accurate short-term solar power forecasting is crucial for grid stability and renewable energy integration. However, solar irradiance is highly variable due to weather conditions, making it challenging for a single forecasting model to perform well across all scenarios. Existing methods often rely on predefined meteorological categories, which may not be readily available or sufficiently granular. This paper aims to address this by developing an unsupervised method to automatically discover and leverage distinct solar irradiance patterns for more adaptive and accurate forecasting.
Methods
The proposed methodology has three core components:
- Optimized Cross-validated ClUsteRing (OCCUR): This novel unsupervised clustering method is designed to determine the optimal number of clusters and the best clustering results for daily Global Horizontal Irradiance (GHI) time series. It achieves this by employing multiple UC algorithms (K-means, K-medoids, AHC, DHC) and cross-validating their performance using internal validity indices (Connectivity, Silhouette width, Dunn's index).
- Support Vector Machine Pattern Recognition (SVM-PR): Once the data is clustered, SVM-PR is used to classify a given day into one of these clusters. Crucially, this classification is performed using only the first few hours of data (specifically, 4 hours in this study), making it practical for real-time forecasting. This approach leverages 13 solar features, including historical GHI, clear sky GHI, sky image statistics (mean, standard deviation, Rényi entropy of nRBR), and meteorological measurements.
- Machine Learning-based Multi-Model (M3) Forecasting Framework: For each identified cluster, a specialized forecasting model is trained. The M3 framework is a two-layer system: the first layer uses various ML algorithms (ANN, SVR, GBM, RF) with different kernels to generate initial forecasts, and the second layer blends these forecasts to produce the final output. This allows the system to select the most suitable M3 model for each cluster, thereby adapting to different solar irradiance patterns.
Datasets
The study utilizes a 1-year hourly GHI and sky imaging dataset collected from the National Renewable Energy Laboratory (NREL) in Colorado. The data was collected at a specific location with known latitude, longitude, and elevation. The dataset was partitioned into training (3:1 ratio) and testing sets, with early morning and late night data (where GHI is near zero) excluded.
Key Insights and Ideas
- Unsupervised Discovery of Patterns: The OCCUR method demonstrates that distinct daily solar irradiance patterns can be automatically discovered without relying on predefined meteorological conditions. This is a significant advantage over traditional approaches.
- Early-Hour Classification is Sufficient: The SVM-PR's ability to accurately classify a day's solar pattern using just the first 4 hours of data is a key enabler. This makes the clustering-based approach practical for short-term forecasting, where timely predictions are essential.
- Tailored Forecasting Yields Superior Accuracy: By assigning each day to a specific cluster and using a dedicated forecasting model for that cluster, the system can better capture the nuances of different solar irradiance behaviors, leading to improved accuracy.
- M3 Framework Enhances Adaptability: The two-layer M3 framework allows for flexible model blending, further optimizing forecasts within each cluster.
Key Results
- Significant Accuracy Improvement: The proposed UC-M3 method outperforms non-UC (all-in-one) models with the same M3 architecture by approximately 20% in terms of normalized mean absolute error (nMAE) and normalized root mean square error (nRMSE).
- M3 Superiority: The M3 framework itself shows a significant improvement (also around 20%) over single-algorithm machine learning (SAML) models.
- High Classification Accuracy: The SVM-PR achieved an overall accuracy of 82.1% for classifying days into clusters using only the first 4 hours of data. This is a substantial improvement over direct classification methods, which required over 11 hours to achieve similar accuracy.
- Calendar and Weather Effects: The study found that forecasting errors exhibit daily and yearly patterns influenced by calendar effects (month and hour) and weather conditions (cluster type). The UC-M3 method generally showed robust and consistent improvements across these variations.
Limitations and Future Work
- Computational Cost: The paper notes that UC-based models and M3-based models require more training time than AIO or SAML models, primarily due to the increased number of forecasting models and the two-layer architecture.
- Model Deterioration in Specific Cases: While generally improving accuracy, the study observed instances where a specific model (e.g., SVR2) showed deteriorated accuracy in one cluster (cluster C) when adopting the UC-M3 approach. However, this deterioration was compensated by significant improvements in other clusters.
- Future Work: The paper suggests that further exploration of different blending algorithms within the M3 framework and investigating more advanced UC algorithms could lead to even greater improvements. The authors also highlight the potential for further research into calendar and weather effects on forecasting accuracy.
Summary updated