Introduction
Short-term load forecasting (STLF) is a cornerstone of modern power grid management, essential for ensuring reliability and cost-efficiency. However, electrical load signals are notoriously difficult to predict due to the complex interplay of resistive and inductive loads, multi-level seasonality (from daily cycles to annual shifts), and non-linear responses to external volatility like extreme weather events. Traditional statistical models and standard Recurrent Neural Networks (RNNs) often fail to capture these long-term dependencies or the intricate relationships between weather and diverse electrical equipment.
Methodology
NeuCast addresses these challenges through a specialized three-pillar architecture that integrates deep learning with classical statistical insights.
Factor Modeling and Seasonal Smoothing
Instead of directly predicting raw power values, NeuCast employs Multi-Layer Perceptrons (MLPs) to map consumption into a latent "hidden space." This approach learns representations for temporal factors (days and hours), allowing the model to capture the co-evolving "signatures" of different electrical loads. To maintain the grid's inherent rhythm, classical smoothing techniques—such as Holt-Winters or seasonal ARIMA—are applied to these latent factors, ensuring the forecast respects established periodicities.
High-level Pattern Purification
A key innovation in NeuCast is "pattern purification." Using the Minimum Description Length (MDL) principle and the AutoPlait algorithm, the framework segments historical data into distinct regimes, such as heatwaves or holidays. By filtering out irrelevant historical noise and focusing on patterns pertinent to the forecast period, the model achieves significantly higher robustness.
Experimental Evaluation
The model was evaluated on two distinct datasets:
- SC Data: A large-scale dataset comprising 133 locations in a major Chinese city over 323 days.
- CMU Data: A 23-day public dataset from Carnegie Mellon University, used to assess performance on shorter time series.
Results and Discussion
NeuCast demonstrated superior performance over state-of-the-art baselines, including GBRT, SVR, and PowerCast. On the SC dataset, the model achieved an 18.1% improvement in RMSE and a 26.9% improvement in Dynamic Time Warping (DTW) distance.
Crucially, NeuCast exhibits remarkable stability. While baseline models often experience a sharp increase in error as the forecast horizon extends, NeuCast maintains a relatively flat error growth curve over a 25-day period, as shown in the comparative error analysis.
Conclusion
NeuCast represents a significant advancement in STLF by combining the non-linear feature extraction capabilities of neural networks with the structural discipline of seasonal smoothing and pattern purification. Future work will explore incorporating spatial relationships between grid locations and extending the framework to other domains like traffic flow and cloud computing load management.
Summary updated