TL;DR
This paper introduces a novel short-term photovoltaic (PV) power forecasting model, the FRD-TLEL, that combines K-means clustering for weather categorization, a two-layer ensemble learning (TLEL) approach with XGBoost, RF, CatBoost, and LSTM, and a feature rise-dimensional (FRD) technique. This method significantly improves both deterministic and probabilistic PV power forecasting accuracy, addressing the challenges posed by PV intermittency for grid stability.
What Problem Does This Paper Address?
The increasing integration of photovoltaic (PV) power into electricity grids offers significant environmental and economic benefits, but its inherent intermittency and volatility pose challenges for grid planning and stable operation. Accurate short-term PV power forecasting is crucial for grid operators to manage supply and demand effectively, optimize resource allocation, and ensure grid reliability. Existing forecasting methods often struggle with the complex, non-linear relationships between weather conditions and PV output, leading to suboptimal accuracy, especially under varying weather patterns. This paper aims to enhance forecasting accuracy and provide probabilistic forecasts to better quantify uncertainty.
Methods
The proposed approach integrates several key techniques:
- K-means Clustering: Used to categorize weather patterns (sunny, cloudy/rainy) based on meteorological data (temperature, humidity, solar irradiance). This allows for tailored forecasting models for different weather conditions.
- Two-Layer Ensemble Learning (TLEL): A sophisticated ensemble method that combines multiple base learners (XGBoost, RF, CatBoost, LSTM). The first layer trains these base models, and their outputs form the input for a meta-learner (LSTM) in the second layer. This leverages the strengths of diverse models and mitigates individual model limitations.
- Feature Rise-Dimensional (FRD) Approach: This novel technique enhances the TLEL model by using the training results of individual models (XGBoost, RF, CatBoost) as new "gain features" for further training. This approach aims to preserve original feature values while achieving a deep fusion of model insights, improving accuracy and efficiency.
- Reciprocal Error Method: Used to combine the forecasting results from the TLEL model and the FRD-optimized models (R-XGBL, R-RFL, R-CatBL). This method amplifies the contribution of models with smaller errors, effectively reducing the overall forecasting error.
- Quantile Regression (QR): Employed for probabilistic forecasting. QR estimates different quantiles of the PV power output, allowing for the generation of prediction intervals at various confidence levels (e.g., 95%, 75%, 50%), thereby quantifying forecast uncertainty.
Datasets
The study utilized a dataset from the Desert Knowledge Australia Solar Center (DKASC) in Alice Springs, Australia. The data spans from April 1, 2016, to March 31, 2018, and includes 15-minute interval measurements of:
- PV power output (kW)
- Solar irradiance (W/m²)
- Ambient temperature (°C)
- Relative humidity (%)
- Wind speed (m/s)
The data was preprocessed to handle missing values and outliers, and the daily forecasting period was restricted to 7:00 to 18:00. A total of 31,725 data points were used.
Key Insights and Ideas
- Weather-Specific Forecasting is Crucial: Clustering weather patterns and training models accordingly significantly improves forecasting accuracy, especially given the distinct behavior of PV output under sunny versus cloudy/rainy conditions.
- Ensemble Learning Outperforms Single Models: Combining multiple diverse models (XGBoost, RF, CatBoost, LSTM) through a two-layer ensemble structure effectively captures complex relationships and improves generalization.
- Feature Engineering with Model Outputs (FRD): The novel FRD approach, which uses model predictions as new features, demonstrates a powerful way to enhance model performance by creating a "deep fusion" of insights from different models.
- Error-Weighted Combination is Key: The reciprocal error method for combining forecasts is effective in mitigating the impact of outliers or poor predictions from individual models, leading to a more robust and accurate final forecast.
- Probabilistic Forecasts Enhance Grid Management: Quantile regression provides valuable uncertainty estimates, allowing grid operators to better plan for potential deviations and ensure grid stability.
Key Results
- Deterministic Forecasting: The proposed FRD-TLEL model achieved the lowest Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) across various seasons and weather types compared to single models (BP, XGBoost, RF, CatBoost, LSTM) and the base TLEL model. For instance, on a sunny winter day, the FRD-TLEL model achieved a MAPE of 1.03% and an RMSE of 1.04 kW, significantly outperforming other models.
- Probabilistic Forecasting: The QR-based interval forecasting demonstrated good performance across different confidence levels (95%, 75%, 50%), seasons, and weather types, indicating reliable prediction intervals.
- Efficiency: The paper also notes that the FRD-TLEL model offers higher forecasting efficiency compared to other advanced models, with significantly shorter average running times.
Limitations and Future Work
- Short-Term Focus: The study primarily focuses on short-term PV power forecasting. The authors suggest exploring models with high forecasting accuracy and efficiency at different time scales as a future research direction.
- Data Specificity: While the DKASC dataset is comprehensive, future work could involve validating the model on datasets from diverse geographical locations and under a wider range of climatic conditions.
- Model Complexity: Although the FRD-TLEL model shows improved accuracy, the ensemble nature and feature engineering steps might increase computational complexity compared to simpler models. Further optimization for real-time deployment could be explored.
Summary updated