Probabilistic Load Forecasting Based on Adaptive Online Learning

Verónica Álvarez, Santiago Mazuelas, José Antonio LozanoIEEE Transactions on Power Systems 2020 · DOI · OpenReview

Abstract

Load forecasting is crucial for multiple energy management tasks such as scheduling generation capacity, planning supply and demand, and minimizing energy trade costs. Such relevance has increased even more in recent years due to the integration of renewable energies, electric cars, and microgrids. Conventional load forecasting techniques obtain single-value load forecasts by exploiting consumption patterns of past load demand. However, such techniques cannot assess intrinsic uncertainties in load demand, and cannot capture dynamic changes in consumption patterns. To address these problems, this paper presents a method for probabilistic load forecasting based on the adaptive online learning of hidden Markov models. We propose learning and forecasting techniques with theoretical guarantees, and experimentally assess their performance in multiple scenarios. In particular, we develop adaptive online learning techniques that update model parameters recursively, and sequential prediction techniques that obtain probabilistic forecasts using the most recent parameters. The performance of the method is evaluated using multiple datasets corresponding with regions that have different sizes and display assorted time-varying consumption patterns. The results show that the proposed method can significantly improve the performance of existing techniques for a wide range of scenarios.

The Challenge of Modern Power Grids

Accurate load forecasting is the backbone of power grid management, affecting everything from generation scheduling to energy trading. However, two major hurdles remain in the current landscape:

  1. Intrinsic Uncertainty: Most conventional methods provide a single-value (point) forecast. In a modern grid with volatile renewable sources and EV charging, knowing the range of possible demand (probabilistic forecasting) is essential for risk management.
  2. Dynamic Patterns: Traditional models are often trained "offline" on historical data. If consumption patterns shift—due to seasonal changes, new infrastructure, or social behavior—these static models become obsolete.

Methodology: Adaptive Probabilistic Load Forecasting (APLF)

The proposed framework uses non-homogeneous Hidden Markov Models (HMMs) to model the relationship between power loads and observations like weather and time. Unlike standard HMMs that require computationally expensive retraining, this method introduces several core innovations:

  • Recursive Parameter Updates: The authors derived closed-form recursive equations to update the model’s means and standard deviations. This allows the model to update itself as every new data point arrives without full retraining.
  • Adaptive Forgetting Factor ($\lambda$): A weighting mechanism prioritizes recent data over old data, allowing the system to "forget" outdated consumption patterns and adapt to new ones in real-time.
  • Sequential Probabilistic Prediction: Instead of a single point, the framework provides a Gaussian density function for future loads. This allows operators to visualize the "confidence" of the forecast through intervals.
Probabilistic load forecast showing confidence intervals around the predicted mean.

Experimental Validation and Results

The APLF method was validated against 11 state-of-the-art techniques using seven publicly available datasets, ranging from large-scale national grids (Belgium, New England) to small-scale building clusters (Australia).

Adaptation to Shifting Patterns

A key insight from the study is the model's ability to "harness" changes in consumption patterns. While offline models continue to predict based on old patterns after a shift, the APLF model adapts within a few cycles, significantly reducing error during transitions.

Comparison between offline and online learning predictions as load patterns shift.

Performance Metrics

  • Superior Accuracy: APLF consistently achieved the lowest Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). In the Belgium dataset, APLF achieved a MAPE of 2.3%, significantly outperforming Gaussian Processes (4.0%).
  • Better Calibration: The model showed superior calibration in its probabilistic forecasts, meaning that predicted confidence intervals accurately reflected the frequency of actual load values falling within them.
  • Robustness at Small Scales: The model proved particularly robust for small-scale regions (100-400 buildings) where load behavior is typically more "jittery" and difficult to model with traditional statistics.

Conclusion

Adaptive Probabilistic Load Forecasting (APLF) addresses the limitations of static, point-based forecasting by combining the flexibility of online learning with the risk-assessment capabilities of probabilistic modeling. By prioritizing recent data and updating parameters recursively, the model remains accurate even as grid dynamics evolve, making it a powerful tool for modern energy management.

Summary updated

Abstract

Load forecasting is crucial for multiple energy management tasks such as scheduling generation capacity, planning supply and demand, and minimizing energy trade costs. Such relevance has increased even more in recent years due to the integration of renewable energies, electric cars, and microgrids. Conventional load forecasting techniques obtain single-value load forecasts by exploiting consumption patterns of past load demand. However, such techniques cannot assess intrinsic uncertainties in load demand, and cannot capture dynamic changes in consumption patterns. To address these problems, this paper presents a method for probabilistic load forecasting based on the adaptive online learning of hidden Markov models. We propose learning and forecasting techniques with theoretical guarantees, and experimentally assess their performance in multiple scenarios. In particular, we develop adaptive online learning techniques that update model parameters recursively, and sequential prediction techniques that obtain probabilistic forecasts using the most recent parameters. The performance of the method is evaluated using multiple datasets corresponding with regions that have different sizes and display assorted time-varying consumption patterns. The results show that the proposed method can significantly improve the performance of existing techniques for a wide range of scenarios.
BibTeX
@article{alvarez2020probabilistic,
  author = {Verónica Álvarez and Santiago Mazuelas and José Antonio Lozano},
  title = {{Probabilistic Load Forecasting Based on Adaptive Online Learning}},
  journal = {IEEE Transactions on Power Systems},
  year = {2020},
  doi = {10.1109/tpwrs.2021.3050837},
}
Figure