TL;DR
This paper introduces a novel data-driven forecasting method for "almost-periodic" time series, like electrical grid load. It combines Dynamic Mode Decomposition (DMD) to capture the deterministic, recurring patterns with Gaussian Process Regression (GPR) to model the unpredictable, stochastic "forcing" that perturbs these patterns. This hybrid approach significantly improves forecasting accuracy compared to existing methods.
What Problem Does This Paper Address?
Accurate time series forecasting is crucial across many scientific disciplines, but it's particularly challenging for complex systems like electrical grids. Grid load exhibits a predictable, "almost-periodic" behavior (daily, weekly cycles) but is also influenced by numerous unpredictable factors (weather, economic shifts, human behavior). Traditional forecasting methods often struggle to balance capturing these regularities with accounting for the inherent randomness. Existing techniques either rely heavily on external explanatory variables (like temperature), which aren't always available or sufficient, or use purely statistical/machine learning models that can be less interpretable. This paper aims to bridge this gap by developing a method that explicitly models both the deterministic and stochastic components of the system's dynamics.
Methods
The core of the proposed method is a fusion of two powerful techniques:
- Dynamic Mode Decomposition (DMD) with Time Delay Embedding: DMD is used to find a linear dynamical system that best describes the observed time series. To handle the "almost-periodic" nature of the data, the time series is first transformed into a higher-dimensional "delay coordinate space" by stacking time-shifted versions of the data. This allows DMD to better capture the underlying linear dynamics. The method specifically uses an "optimized DMD" (optDMD) variant that can be constrained to produce purely imaginary eigenvalues, reflecting stable oscillations without growth or decay.
- Stochastic Forcing with Gaussian Process Regression (GPR): After DMD identifies the deterministic linear dynamics, the discrepancies between the model's predictions and the actual data are treated as a "forcing" signal. This forcing represents the unpredictable, non-periodic influences on the system. GPR is then employed to learn a probabilistic model of this forcing signal. This allows for forecasting by simulating the deterministic DMD dynamics and then adding a stochastically sampled forcing term at each time step.
This approach effectively separates the predictable, quasi-periodic components of the system from the unpredictable, stochastic ones, modeling each with a suitable mathematical tool.
Datasets
The primary dataset used for training and evaluation is hourly electrical load data from the Duke Energy power company, covering the states of North Carolina and South Carolina for a full year (January 1, 2017, to December 30, 2017). The models were trained on 4-week subsets of this data.
Key Insights and Ideas
- "Almost-Periodic" Systems as Forced Linear Systems: The key insight is that many complex real-world time series, like grid load, can be effectively modeled as a deterministic linear system that is perturbed by a stochastic forcing. This decomposition allows for a more nuanced and accurate representation.
- Time Delay Embedding for DMD: Applying DMD in a time-delay coordinate space is crucial for capturing the quasi-periodic nature of the data, making the underlying dynamics more amenable to linear modeling.
- Separating Deterministic and Stochastic Dynamics: The method's strength lies in its ability to disentangle the predictable, cyclical behavior (modeled by DMD) from the unpredictable, external influences (modeled by GPR). This separation leads to more robust forecasts.
- Interpretability through Forcing: The learned forcing signal offers a novel avenue for understanding anomalous events. Large spikes in the forcing can often be correlated with holidays, extreme weather, or other unusual occurrences, providing diagnostic insights.
Key Results
- Superior Forecasting Accuracy: The proposed stochastically forced DMD method consistently outperforms state-of-the-art forecasting techniques (including LSTM, ARIMA, GPR, and ensemble GPR) on the Duke Energy load data, especially for forecasts extending up to two weeks.
- Competitive Performance Even Without Forcing: Even the "unforced" DMD component of the model shows competitive performance against other leading methods.
- Improved Phase Accuracy: The stochastic forcing component helps correct phase errors that accumulate in unforced DMD forecasts over time.
- Interpretability: The linear nature of the DMD component allows for spectral analysis, and the learned forcing signal provides insights into system anomalies.
Limitations and Future Work
- Assumption of "Almost-Periodic" Dynamics: The method is most effective for systems with a continuous Fourier spectrum punctuated by dominant peaks. Systems with truly continuous spectra might not be well-represented.
- Forcing Model Complexity: While GPR is used for the forcing, the paper notes that estimating the forcing signal using multiple independent GPRs might be less effective than methods accounting for correlations among outputs.
- Interpretability of Forcing: While the forcing signal shows correlations with holidays, further work is suggested to better understand its relationship with other environmental factors like temperature, precipitation, and economic trends. This could lead to improved stochastic modeling of the forcing.
- Uncertainty Quantification: The framework naturally lends itself to uncertainty quantification by sampling ensembles of the stochastic forcing. Future work could explore how to tune this to better match observed dynamics.
- Generalizability: The authors believe the method can generalize to other complex, noisy, or stochastic time series beyond power grid load.
Summary updated