Introduction
As the solar photovoltaic (PV) industry matures into a multi-billion dollar sector, the focus of operational management is shifting from individual sites to massive fleets comprising thousands of systems. This transition introduces significant data challenges. Traditional fleet modeling often relies on bespoke, labor-intensive models for each site, which fail to scale efficiently or capture the complex interdependencies between systems.
The primary hurdles in modeling PV fleets include the strict diurnal periodicity of solar output, the spatiotemporal correlations caused by moving weather patterns (e.g., cloud cover), and the inherent noise or missing values in real-world sensor data. This paper introduces a scalable, "white-box" probabilistic framework designed to model the joint probability distribution of an entire PV fleet across both space and time using a smooth periodic Gaussian copula.
Methodology: The Gaussianization Pipeline
The authors propose a three-step invertible pipeline to transform raw power data ($y_t$) into a "whitened" signal ($z_t$) that follows a standard Normal distribution $N(0, I)$. This process, known as Gaussianization, allows complex joint distributions to be handled using the well-understood tools of Gaussian statistics.
1. Smooth Periodic Marginal Transform
PV power data is non-Gaussian; it is bounded by zero and a maximum "clear-sky" potential. To address this, the authors employ periodic quantile regression. By representing quantiles with Fourier series and minimizing a combination of "pinball loss" and a Dirichlet energy smoothness penalty, they create a time-varying map. This map transforms the marginal distribution of the data at any given time into a standard Gaussian distribution.
2. Vector Autoregressive (AR) Model
To capture temporal dependencies—such as the high correlation between power output at 12:00 PM and 12:15 PM—the framework fits a vector AR model. This step accounts for how the current state of the entire fleet depends on its recent historical values, effectively "de-correlating" the data in the time domain.
3. Smooth Periodic Residual Fit
The residuals remaining after the AR model still exhibit correlations between different systems that vary according to the time of day. The authors fit a time-varying Gaussian distribution to these residuals. By using Cholesky factors within the optimization, they ensure the loss function remains convex, guaranteeing a global optimum and efficient computation.
Experimental Setup and Data
The model was validated using 15-minute average power data provided by SunPower Corporation. The dataset focused on six PV systems located in Southern California (Santa Ana and Tustin). The training period spanned March 2017, encompassing both clear-sky and volatile, cloudy conditions. A subsequent two-week period in April was used for validation.
Key Insights and Applications
The "white-box" nature of this framework offers several advantages over "black-box" deep learning approaches like LSTMs:
- Interpretability: Operators can directly inspect time-varying means, standard deviations, and inter-system correlations. This transparency is crucial for understanding fleet behavior under varying atmospheric conditions.
- Asymmetric Uncertainty: The model naturally captures the asymmetric nature of solar risk. On clear days, the "downside" risk (potential cloud cover) is significantly larger than the "upside" potential, a nuance captured by the nonlinear marginal transforms.
- Robust Data Imputation: Because the model defines a joint Gaussian distribution, missing data points can be imputed using conditional expectations—simple linear algebra—rather than iterative or heuristic schemes.
- Advanced Anomaly Detection: The framework identifies anomalies by flagging data points that fall into extremely low-probability conditional quantiles. This allows the model to distinguish between actual system faults and expected variations due to heavy cloud cover.
Results
The framework demonstrated high fidelity in generating synthetic power curves that are statistically indistinguishable from real-world data. In anomaly detection tests, the model successfully identified 100% of artificial 15% perturbations. Furthermore, the model provides probabilistic forecasting with 10th and 90th percentile confidence bands that appropriately widen during volatile weather and tighten during clear-sky conditions.
Conclusion and Future Directions
While the current study demonstrates the framework on a fleet of six systems, the authors highlight its potential for broader scaling. Future work will focus on the computational challenges of scaling the joint Gaussian fit to thousands of systems and the integration of exogenous variables, such as satellite-derived irradiance and cloud cover maps, to further refine forecasting accuracy.
Summary updated