Abstract
Load forecasting is an integral part of power system operations and planning. Due to the increasing penetration of rooftop PV, electric vehicles and demand response applications, forecasting the load of individual and a small group of households is becoming increasingly important. Forecasting the load accurately, however, is considerable more difficult when only a few households are included. A way to mitigate this challenge is to provide a set of scenarios instead of one point forecast, so an operator or utility can consider a range of behaviors. This paper proposes a novel scenario forecasting approach for residential load using flow-based conditional generative models. Compared to existing scenario forecasting methods, our approach can generate scenarios that are not only able to infer possible future realizations of residential load from the observed historical data but also realistic enough to cover a wide range of behaviors. Particularly, the flow-based models utilize a flow of reversible transformations to maximize the value of conditional density function of future load given the past observations. In order to better capture the complex temporal dependency of the forecasted future load on the condition, we extend the structure design for the reversible transformations in flow-based conditional generative models by strengthening the coupling between the output and the conditional input in the transformations. The simulation results show the flow-based designs outperform existing methods in scenario forecasting for residential load by both providing more accurate and more diverse scenarios.
Introduction
Forecasting electricity demand for individual households or small clusters is a significant challenge in modern power systems. Unlike aggregated city-wide loads, which exhibit smooth and predictable patterns, residential loads are highly volatile. This volatility stems from the stochastic nature of human behavior—activities like charging an electric vehicle or using high-power appliances occur at irregular intervals.
Traditional "point forecasts" provide a single expected value for future time steps but fail to capture the inherent uncertainty of these volatile loads. To address this, scenario forecasting aims to generate multiple "plausible futures." This allows utility operators to perform robust planning by considering a range of potential outcomes rather than relying on a single, often inaccurate, estimate.
Methodology: Flow-Based Generative Models
The proposed approach leverages Conditional Normalizing Flows, specifically an architecture inspired by Glow. Normalizing flows learn to map complex data distributions (like residential load profiles) to a simple standard Gaussian distribution through a sequence of reversible transformations.
Reinforced-flow Architecture
A key innovation in this work is the Reinforced-flow design. In standard "vanilla" conditional flows, only a portion of the data is typically transformed based on the conditional input (the historical load). The reinforced architecture ensures a tighter coupling between the historical context and the generated scenarios. This allows the model to better capture how past usage patterns influence future behavior, such as the timing of peak demand.
Wasserstein Regularization
Standard flow-based models are typically trained using Maximum Likelihood Estimation (MLE), which minimizes KL-divergence. However, MLE can lead to "over-smoothed" distributions that cover all possible data points with high variance. To combat this, the authors introduce a weighted Wasserstein distance metric into the training objective. This regularization helps the model produce "sharper" and more physically realistic scenarios that concentrate around the most likely realizations without sacrificing necessary diversity.
Experimental Evaluation
The models were evaluated using the Dataport dataset from Pecan Street Corporation, featuring hourly power consumption from 105 households in Austin, Texas, over five years (2013–2017). The study compared performance across different aggregation levels: single households, 10 households, and 100 households.
Key Results
The flow-based models, particularly the reinforced version, demonstrated superior performance in terms of both reliability and sharpness.
- Reliability: Deviation-Coverage Area plots indicated that the generated scenarios were more statistically consistent with actual realizations compared to traditional autoregressive models.
- Sharpness: The 50% prediction intervals (PI) were adaptive; they remained narrow during predictable base-load periods and widened appropriately during volatile peak hours.
- Visual Realism: Qualitatively, the generated scenarios captured the "spiky" nature of real residential demand, which is often lost in smoothed point forecasts.
Conclusion
This research presents a robust framework for residential load scenario forecasting using reinforced conditional normalizing flows. By strengthening the coupling between historical data and future projections and incorporating Wasserstein regularization, the model produces diverse yet realistic scenarios. Future work will explore the integration of exogenous variables, such as weather data and occupancy sensors, to further refine the conditioning process and handle highly noisy historical data.
Summary updated