TL;DR
This paper introduces a novel privacy-preserving federated learning approach for ultra-short-term photovoltaic (PV) power forecasting. By combining collaborative feature learning with a joint probability distribution model, it significantly improves forecasting accuracy and quantifies uncertainty while safeguarding sensitive data from distributed PV systems.
What Problem Does This Paper Address?
Accurately forecasting the power output of distributed photovoltaic (DPV) systems is crucial for grid stability and market operations. However, DPV systems are inherently intermittent and random, making forecasting challenging. Existing methods often struggle with a trade-off: either they require extensive data sharing, risking privacy breaches, or they limit information exchange, leading to suboptimal forecasting accuracy. Furthermore, most current methods focus on deterministic point forecasts, failing to adequately capture the inherent uncertainty in PV generation. This paper aims to bridge this gap by developing a method that enhances forecasting accuracy through collaborative learning while preserving data privacy and providing probabilistic forecasts to quantify uncertainty.
Methods
The proposed method has two main phases: point forecasting and probabilistic forecasting.
Point Forecasting: * Collaborative Feature Federated Learning: This framework addresses privacy concerns by avoiding direct raw data sharing. * Central Server: Aggregates global models based on client capacity and uses Independent Component Analysis (ICA) to generate "global features" from local temporal features, capturing spatial information across DPV sites. * Client Sites: Employ a Transformer-based autoencoder to extract local temporal features. These local features are then fused with the global features from the central server to create "spatiotemporal correlation features." An attention mechanism helps the model focus on critical features for forecasting.
Probabilistic Forecasting: * Joint Probability Distribution Modeling: This approach moves beyond point forecasts by modeling the relationship between forecast values and forecasting errors. * A Copula function is used to model the joint distribution of point forecast values and their corresponding errors. This allows for a more accurate quantification of uncertainty by considering the dependencies between forecast outcomes and error distributions. * The method then derives forecasting intervals by analyzing the conditional probability distribution of errors based on the point forecast.
Datasets
The experimental data consists of power generation data from 30 DPV stations in northern China, spanning from July 1, 2023, to July 1, 2024, with a 15-minute resolution. The data was cleaned and preprocessed, and then split into training, validation, and test sets (8:1:1 ratio).
Key Insights and Ideas
- Privacy-Preserving Collaboration is Key: The paper demonstrates that effective collaboration among DPV sites is possible without compromising data privacy, a significant hurdle in distributed systems.
- Global Features Enhance Local Models: Generating "global features" at the central server and distributing them to clients provides valuable context that local models alone cannot capture, leading to improved forecasting.
- Spatiotemporal Correlations are Crucial: By explicitly modeling the relationships between different DPV stations (spatiotemporal correlations), the accuracy of individual station forecasts is significantly boosted.
- Uncertainty Quantification Matters: Moving from deterministic to probabilistic forecasts, by modeling the joint distribution of forecast values and errors, provides a more complete picture of future power output and is essential for robust grid management.
- Error Distribution is Not Uniform: The study highlights that forecasting errors are not independent and their distribution is closely tied to the PV output's fluctuation characteristics, a key insight for building more accurate probabilistic models.
Key Results
- Point Forecasting Accuracy: The proposed method achieved superior point forecasting performance. Compared to Method 1 and Method 2, it improved NRMSE by up to 2.59% and ACC by up to 2.86% at the 4-hour ahead time scale.
- Probabilistic Forecasting Performance: At a 90% confidence level, the proposed method achieved the best Skill Score (SS), indicating a good balance between interval coverage (PICP) and interval width (PINAW). For example, it improved SS by over 21% compared to Method 4.
- Robustness Across Weather Conditions: The method demonstrated strong performance across sunny, cloudy, and overcast conditions, outperforming other methods, especially in challenging overcast and cloudy scenarios where irradiance fluctuations are high.
- Transformer Autoencoder Efficacy: The Transformer autoencoder proved effective in capturing local temporal dependencies and focusing on critical features, outperforming a GRU model.
Limitations and Future Work
- Model Parameter Aggregation: While global features are aggregated, the paper suggests future work could explore more advanced methods for aggregating model parameters themselves to further enhance federated learning performance.
- Copula Function Variety: Only the Clayton Copula function was used. Future research could investigate other Copula functions to better fit PV data and improve probabilistic forecasting.
- Weather Anomaly Detection: The authors propose incorporating a weather anomaly detection mechanism to assess model performance under extreme weather events and grid instability conditions.
Summary updated