Complex-Valued Time Series Based Solar Irradiance Forecast

This paper proposes a resource-efficient complex-valued autoregressive model that transforms solar irradiance forecasts by incorporating volatility as an imaginary component, demonstrating superior or comparable accuracy to classical methods while offering broad applicability across various scientific fields.

The Gist

The Big Picture: Predicting the Sun's Mood

Imagine you are a chef trying to bake a perfect cake, but your oven is powered entirely by the sun. The problem? The sun is moody. Sometimes it shines brightly, and sometimes a cloud zips by, causing the power to flicker wildly.

To keep your cake from burning or undercooking, you need to know not just how much sun you'll get tomorrow, but also how likely it is to change suddenly.

This paper introduces a new, clever way to predict solar energy. Instead of just guessing the amount of sunlight, it predicts the sunlight and the sunlight's mood swings (volatility) at the same time, using a mathematical trick involving "complex numbers."

---

The Problem: Why Old Methods Fail

Most current weather apps for solar power are like a slow-moving turtle.

• They are great at predicting the general trend (e.g., "It will be sunny in the afternoon").
• But they are terrible at predicting sudden changes (e.g., "A cloud is passing right now, and the power will drop in 30 seconds").

Scientists have tried to build "super-fast" models to catch these changes, but those models are often too complicated, require too much data, and crash easily. Grid managers (the people running the power grid) prefer simple, robust tools, even if they aren't perfect.

The Solution: The "Complex" Magic Trick

The authors propose a new method that is simple to build but surprisingly smart. Here is how it works, step-by-step:

1. The Two-Part Recipe (Real vs. Imaginary)

In math, a "complex number" has two parts: a Real part and an Imaginary part.

• The Real Part: This is the actual sunlight measurement (how bright it is right now).
• The Imaginary Part: This is the moodiness (how much the light is flickering or changing).

Usually, people treat these as two separate problems. This paper says: "Let's mix them together into one single package."

The Analogy: Imagine you are driving a car.

• The Real part is your speedometer (how fast you are going).
• The Imaginary part is your shakiness (how much the road is bumping).
• Instead of looking at the speedometer and the road separately, this new method looks at a single "Drive-O-Meter" that tells you both your speed and how bumpy the ride is about to be, all in one glance.

2. The "Crystal Ball" (The Model)

The researchers use a simple mathematical tool called an Autoregressive (AR) model. Think of this as a smart mirror.

• It looks at the last 30 hours of data (the "Real" speed and the "Imaginary" bumps).
• It uses a simple formula to guess what the next hour will look like.
• Because it treats the "bumps" (volatility) as part of the same equation as the "speed" (sunlight), it can react much faster to sudden changes than traditional models.

3. The Safety Net (Prediction Intervals)

The most useful part of this paper isn't just the guess; it's the confidence level.

• Old way: "We think it will be 500 Watts." (No idea if it could be 100 or 900).
• New way: "We think it will be 500 Watts, but based on the current 'moodiness,' it will likely stay between 450 and 550 Watts."

This creates a safety band. If the band is narrow, the sun is calm. If the band is wide, the sun is chaotic, and the power grid needs to be ready for a shock.

Why This is a Big Deal

The authors tested this on data from Corsica (an island in France with lots of sun and clouds). Here is what they found:

1. It's Simple: You don't need a supercomputer. You could literally do this calculation in a basic spreadsheet.
2. It's Accurate: It predicts the sunlight just as well as the heavy-duty, complicated models.
3. It's Better at Safety: When it comes to predicting the "safety band" (the range where the sun might actually be), this new method is tighter and more accurate than the old methods. It gives grid managers a better "safety net."

The Bottom Line

Think of this new method as upgrading from a static map to a live GPS.

• The Static Map (old models) tells you the road is straight.
• The Live GPS (this new method) tells you the road is straight right now, but warns you that there is a pothole coming up in 5 minutes, so you should slow down.

By treating the "flicker" of the sun as a mathematical partner to the "brightness," the researchers created a tool that is cheap, fast, and incredibly useful for keeping our solar-powered world running smoothly.

Technical Summary

1. Problem Statement

The integration of solar energy into power grids is hindered by the random and variable nature of solar irradiance. While numerous machine learning methods exist for forecasting, they often face a trade-off:

• Simple models are robust and easy to implement but fail to capture rapid fluctuations (cloud transients).
• Advanced non-parametric models capture fluctuations better but are prone to overfitting, require excessive data, or are too computationally complex for practical grid management.
• Existing probabilistic methods (e.g., Gaussian processes, bootstrapping) often produce prediction intervals that are either too wide to be useful or rely on restrictive assumptions (like Gaussian residuals) that do not hold for solar data.

The authors propose a new, lightweight, non-parametric probabilistic forecasting method that simultaneously predicts the Global Horizontal Irradiance (GHI) and its volatility (uncertainty) using a complex-valued time series approach.

2. Methodology

The core innovation is the transformation of the forecasting problem into a complex-valued autoregressive (AR) model.

A. Data Preprocessing

• Input: Raw GHI measurements from Ajaccio, Corsica (Mediterranean climate).
• Clear-Sky Index ($\kappa$): Instead of predicting raw GHI, the model predicts the clear-sky index $\kappa(t) = \text{GHI}(t) / \text{GHICS}(t)$, where $\text{GHICS}$ is the theoretical clear-sky irradiance calculated via the Solis model. This normalizes the data to $[0, 1]$ and removes seasonal/diurnal trends.
• Volatility ($\sigma_\tau$): A volatility series is constructed based on the standard deviation of the returns of $\kappa$ over a sliding window ($\tau = 30$ hours). This represents the "imaginary" part of the complex series, capturing the risk of rapid fluctuations.

B. Complex-Valued Transformation

The authors define a complex scalar time series $z(t)$:
$$z(t) = \kappa(t) + j\sigma_\tau(t)$$
Where:

• Real part ($\Re(z)$): The clear-sky index (trend).
• Imaginary part ($\Im(z)$): The estimated volatility (fluctuation risk).

C. Complex Autoregressive Model (AR)

A standard AR($p$) model is adapted for complex numbers:
$$\hat{z}(t+1) = \sum_{i=0}^{p-1} z(t-i)\omega_i$$

• Model Identification: The order $p$ is determined by analyzing the partial autocorrelation of both real and imaginary parts. The authors set $p = \max(p_\Re, p_\Im)$, typically between 2 and 6.
• Parameter Estimation: The coefficients $\omega$ are estimated using Least Squares optimization extended to the complex domain (using the Frobenius norm). To prevent overfitting and improve the condition number, a Ridge regression approach (L2 regularization) is applied, with the regularization parameter $\lambda$ tuned via cross-validation.

D. Probabilistic Forecasting

Once $\hat{z}(t+1)$ is predicted:

1. Point Forecast: $\hat{\text{GHI}} = \Re(\hat{z}) \times \text{GHICS}$.
2. Prediction Intervals: The imaginary part $\Im(\hat{z})$ provides the volatility estimate. Prediction intervals are constructed assuming a Gaussian distribution of the residuals, bounded by:
   $$\Lambda = [\hat{\text{GHI}} - \mu \cdot \text{GHICS} \cdot \hat{\sigma}_\tau, \quad \hat{\text{GHI}} + \mu \cdot \text{GHICS} \cdot \hat{\sigma}_\tau]$$
   Note: The scaling factor $\mu$ is calibrated using a data-driven approach (simulations on training data) rather than purely theoretical assumptions, allowing the method to remain non-parametric regarding the distribution shape.

3. Key Contributions

• Novel Formalism: Introduction of a complex-valued time series framework where the imaginary part explicitly models volatility, allowing a single AR model to predict both the value and the uncertainty of solar irradiance.
• Simplicity and Efficiency: The method requires very few parameters (only 6 complex numbers in the study) and minimal computational resources, making it suitable for standalone applications and edge computing in smart grids.
• Theoretical Justification: The paper provides a theoretical link between the defined volatility and the variance of high-frequency fluctuations, validating the use of volatility to bound prediction intervals.
• Data-Driven Calibration: The authors move beyond theoretical Gaussian assumptions by calibrating the prediction interval width using empirical data, improving robustness.

4. Results

The model was tested on data from 2008–2017 (training) and 2018 (testing) for horizons of 1 to 6 hours. It was compared against:

• Gaussian Process (Gauss)
• Bootstrap Methodology (Boot)
• Quantile Regression (Quant)

Key Performance Metrics:

• Deterministic Accuracy (nRMSE): The complex model (Compl) achieved the lowest error across all horizons (e.g., 0.196 at 1h vs. 0.197 for Gauss), demonstrating it captures trends as well as complex models.
• Probabilistic Accuracy (PICP): At a nominal 80% coverage, the Compl model achieved 80.01% actual coverage, closely matching the target, whereas the Bootstrap method under-performed (75.21%).
• Interval Sharpness (MIL): The Compl model produced significantly narrower prediction intervals (Mean Interval Length) compared to other methods while maintaining coverage. For example, at 1h, Compl had an MIL of 41.24%, compared to 51.24% for Gauss.
• Overall Score (MSIS): The Compl model achieved a Mean Scaled Interval Score of 0.95, outperforming Bootstrap (1.05) and Gaussian (1.03), and was competitive with Quantile Regression (0.89).

5. Significance and Conclusion

This paper demonstrates that complex-valued time series analysis offers a powerful, yet simple, alternative to sophisticated machine learning for solar forecasting.

• Practicality: It bridges the gap between simple robust models and complex fluctuation-capturing models. Grid operators can implement this using basic tools (e.g., spreadsheets) without heavy computational infrastructure.
• Scalability: The methodology is not limited to solar energy; it can be applied to any univariate time series where trend and volatility need simultaneous prediction (e.g., wind power, electricity load).
• Future Work: The authors suggest extending the imaginary part to include other variables (exogenous data, residuals) and adapting the framework to other predictors like Neural Networks or Support Vector Regression.

In summary, the proposed method successfully delivers high-accuracy deterministic forecasts and tight, reliable probabilistic intervals with minimal computational cost, addressing a critical need for efficient renewable energy management.