TLDR: This paper introduces a global modeling framework for scalable short-term electricity load forecasting in heterogeneous power grids. It addresses challenges like data drift and diverse consumption patterns by proposing two novel time series clustering (TSC) methods: model-based TSC for feature-transforming models and weighted instance-based TSC for target-transforming models. Experiments on Alberta’s electricity load data demonstrate that these globalization and clustering techniques significantly improve forecasting accuracy, scalability, and robustness, especially for peak load and zero-shot hierarchical forecasting.
Forecasting electricity load is crucial for managing power transmission networks efficiently, from large-scale systems down to individual points of delivery. Traditionally, local forecasting models (LFMs) have been used, treating each measuring point as a separate entity. While intuitive, these models face significant challenges, especially when dealing with large datasets. They struggle with scalability, becoming computationally expensive and less efficient as the network grows. LFMs also suffer from issues like overfitting with short time series, the ‘cold start’ problem for new locations with limited historical data, and a lack of generalizability, failing to capture commonalities across related time series.
A new approach, global forecasting models (GFMs), offers a promising alternative. GFMs enhance prediction generalizability, scalability, accuracy, and robustness by leveraging ‘globalization’ and ‘cross-learning’. Instead of training a separate model for each time series, a single global model is trained on combined data from all series, capturing shared patterns. This approach significantly reduces the number of models needed, streamlining training and deployment. GFMs can also predict for time series with little or no data, effectively addressing the cold start problem, and are more robust due to learning common parameters across multiple series.
However, GFMs have a notable drawback: they assume that all input time series are inherently related. This assumption often overlooks significant spatiotemporal data heterogeneity—differences in patterns across various locations or within a single series over time. For example, traditional consumers might become ‘prosumers’ with solar panels, leading to data drift. External events like wildfires or pandemics can also introduce sudden shifts in data distribution. A single GFM might struggle to accurately capture the behavior of diverse time series, potentially sacrificing local accuracy for broader generalization.
To address this challenge and strike a balance between global and local dynamics, this research proposes implementing time series clustering (TSC). The idea is to group time series with shared characteristics, then train a GFM for each distinct cluster. This allows the model to leverage shared patterns within clusters while still accounting for heterogeneity across them.
The paper introduces two distinct TSC approaches tailored to different types of forecasting models:
Model-based Whole Time Series Clustering
This method is designed for ‘feature-transforming’ algorithms, such as linear regression and neural networks, which learn a mathematical function to combine input features. Here, local models are first trained for each time series, and their learned coefficients (e.g., weights from linear regression) are used to cluster series based on process-driven similarities. This ensures that the global model within each cluster effectively captures the underlying patterns common to its members, improving accuracy while balancing locality and globality.
Also Read:
- A New Hybrid AI Model for Smarter Energy Forecasting
- Boosting Time Series Forecasts with Foundation Models and Conformal Prediction
Weighted Instance-based Time Series Clustering
This approach is developed for ‘target-transforming’ algorithms, like decision trees and nearest neighbors, which group and average target values from the training set. In this method, a global model is initially trained, and its coefficients are used to weight the importance of features during the clustering process. This creates a weighted Euclidean distance metric, allowing the clustering algorithm to prioritize the most impactful features and form more meaningful clusters that align with the data’s true underlying patterns.
The proposed framework was extensively tested using a real-world dataset of Alberta’s electricity load, comprising 42 time series from different areas across the province. This dataset exhibits significant spatiotemporal heterogeneity due to diverse consumption types (residential, commercial, industrial), varying climatic conditions, and events like the COVID-19 pandemic and wildfires, which cause data drift.
The experiments demonstrated that global target-transforming models (like LightGBM and XGBoost) consistently outperformed their local counterparts, especially when enriched with global features and clustering techniques. The weighted instance-based clustering specifically improved their performance by retaining diverse instances within clusters. For global feature-transforming models (like Ridge regression), model-based clustering proved highly effective, allowing them to adapt better to localized patterns while maintaining their linear structure.
The research also assessed the impact of globalization on peak load forecasting and hierarchical forecasting. Global models, particularly when combined with appropriate clustering, showed improved accuracy in predicting peak load events, which are critical for grid management. Furthermore, the models demonstrated strong ‘zero-shot’ forecasting capabilities for regional and system-level loads, meaning they could accurately predict for aggregated levels even without being explicitly trained on that specific hierarchical level. This highlights the potential for a unified framework for hierarchical load forecasting across all levels of transmission networks.
In summary, this paper provides a comprehensive analysis of data heterogeneity and drift in electricity load data and proposes a global modeling framework for scalable short-term load forecasting. It introduces tailored time series clustering approaches that effectively balance globality and locality, enhancing forecasting accuracy and robustness in heterogeneous power grids. For more details, you can refer to the full research paper: Globalization for Scalable Short-term Load Forecasting.
