STL Decomposition

Unlike classical decomposition methods that rely on simple moving averages, STL uses Locally Estimated Scatterplot Smoothing (LOESS). This non-parametric method fits low-degree polynomials to localized subsets of the data. This provides superior handling of non-linear trends and irregular seasonal patterns common in solar irradiance data, and is inherently robust to outliers that would otherwise distort the decomposition.

STL decomposes a time series into three additive components:

• Trend (T_t): The long-term, low-frequency variation, such as gradual panel degradation.
• Seasonal (S_t): The repeating, high-frequency pattern, like the diurnal solar cycle.
• Remainder (R_t): The stochastic, high-frequency noise left after extracting trend and seasonality, representing weather-driven volatility. This clean separation allows forecasters to model each component independently.

A defining strength of STL is its ability to handle seasonal components that evolve over time. The seasonal pattern is not assumed to be constant; it can slowly change in amplitude and shape. This is critical for capturing the shifting shape of the daily solar curve across different months or the gradual change in diurnal wind patterns, providing a more accurate baseline than fixed seasonal dummies.

The algorithm operates through two nested iterative loops:

• Inner Loop: Alternates between updating the seasonal and trend components using LOESS smoothing and moving averages to converge on a stable decomposition.
• Outer Loop: Calculates robustness weights based on the magnitude of the remainder component. Data points with large remainders (anomalies) are down-weighted in subsequent inner loop iterations, making the decomposition resilient to cloud-induced irradiance spikes or sensor errors.

The user controls the decomposition's granularity through key parameters:

• n_p: The number of observations per seasonal cycle (e.g., 24 for hourly diurnal data).
• s_window: The span of the LOESS window for seasonal extraction, controlling how rapidly the seasonal pattern can evolve.
• t_window: The span for trend extraction, defining the smoothness of the long-term component. This configurability allows precise tuning for intraday forecasting versus day-ahead trend analysis.

In renewable generation forecasting, STL is rarely the final model but a critical preprocessing step. By removing the deterministic diurnal cycle (seasonal component), the resulting stationary remainder can be fed into machine learning models like Long Short-Term Memory (LSTM) networks or Gradient Boosting Machines. These models then focus solely on learning the complex, weather-driven residual dynamics, significantly improving forecast accuracy.

The core statistical engine driving STL. Loess fits low-degree polynomials to localized subsets of data using weighted least squares, giving more influence to points near the target. In STL, an inner loop applies Loess iteratively to update the seasonal component, while an outer loop uses Loess on the residuals to compute robustness weights, mitigating the influence of outliers on the trend and seasonal estimates.

A versatile and robust filtering procedure for decomposing a time series into trend, seasonal, and remainder components. Unlike classical decomposition, STL handles any seasonal periodicity, allows the seasonal component to evolve over time, and is robust to anomalous observations. It is widely applied to load profiling to separate predictable diurnal human behavior from long-term electrification trends and stochastic weather-driven noise.

A simpler precursor to STL that decomposes a time series into trend, seasonal, and random components using moving averages. It assumes a fixed, non-evolving seasonal pattern and is highly sensitive to outliers. While computationally cheap, it is generally unsuitable for energy data where consumption patterns shift due to demand response programs or seasonal amplitude changes driven by climate variability.

A sophisticated seasonal adjustment software package developed by the U.S. Census Bureau. It combines the RegARIMA modeling framework (for pre-adjusting outliers, calendar effects, and holiday impacts) with signal extraction filters. For grid forecasting, X-13ARIMA-SEATS excels at modeling complex calendar effects—such as moving holidays and daylight saving time transitions—that heavily influence commercial and residential load patterns.

An adaptive, data-driven method for decomposing a non-linear and non-stationary signal into Intrinsic Mode Functions (IMFs) . Unlike STL, EMD does not require pre-specifying a seasonal period, making it useful for analyzing transient grid events like inter-area oscillations or fault-induced ringing that lack a fixed frequency. It is often used as a pre-processing step for Hilbert spectral analysis of power quality disturbances.

A mathematical tool that decomposes a signal into components at different frequency scales, providing simultaneous time-frequency localization. In energy forecasting, DWT is used to denoise price spikes or separate high-frequency renewable ramp events from the smoother, low-frequency diurnal cycle. It provides a multi-resolution analysis that complements STL's additive decomposition structure.