Learning-based Probabilistic Load Forecasting with Post-hoc and In-model Uncertainty

2026-07-14 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

A new unified one-day-ahead probabilistic load forecasting framework addresses challenges arising when smart-building load forecasters, trained on dense, high-frequency data, are deployed with sparse, hourly inputs requiring feature reconstruction. The framework compares a modular post-hoc residual-quantile scheme against an integrated in-model quantile-learning scheme, using recurrent, hybrid recurrent, and Temporal Fusion Transformer (TFT) Deep Learning backbones. Results indicate that optimal uncertainty placement is backbone-dependent. The integrated quantile learning approach proved most reliable with the TFT model, achieving 2.2-3.6% MAPE and 28-83W RMSE on the test window, while producing prediction intervals approximately 5x narrower than modular methods at comparable coverage. A reconstruction-sensitivity test revealed that reconstructed inputs increase the Quantile Score by 106% without a corresponding increase in interval width, demonstrating that models do not inherently absorb reconstruction-induced uncertainty. This highlights limitations of post-hoc residual quantiles when inference relies on reconstructed inputs.

Key takeaway

For Machine Learning Engineers deploying smart-building load forecasting models where inference relies on reconstructed inputs, carefully evaluate your uncertainty quantification strategy. The study shows that integrated in-model quantile learning, particularly with Temporal Fusion Transformer (TFT) backbones, yields significantly narrower and more reliable prediction intervals compared to post-hoc methods. You should explicitly account for input reconstruction uncertainty, as models do not automatically absorb these errors, impacting demand-response scheduling and overall forecast utility.

Key insights

Optimal uncertainty placement in probabilistic load forecasting is backbone-dependent, with integrated quantile learning excelling for TFT models.

Principles

• Input uncertainty from reconstructed features can miscalibrate prediction intervals.
• Models do not automatically absorb reconstruction-induced uncertainty.
• Integrated quantile learning can yield narrower, more reliable intervals.

Method

A unified framework aligns temporal resolution, reconstructs unavailable inputs, derives causal features, then compares post-hoc residual-quantile and in-model quantile-learning schemes.

In practice

• Prioritize integrated quantile learning for TFT-based load forecasting.
• Evaluate uncertainty placement strategy based on chosen DL backbone.

Topics

• Probabilistic Load Forecasting
• Deep Learning
• Temporal Fusion Transformer
• Uncertainty Quantification
• Feature Reconstruction
• Smart Buildings

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Uncertainty-Aware Transfer Learning for Cross-Building Energy Forecasting: Toward Robust and Scalable District-Level Energy Management — Uncertainty-aware transfer learning with TFT enables robust cross-building energy forecasting using minimal target data.
• A Benchmark for Electrical Load Forecasting Across Grid Levels: Time-Series Transformers Outperform Established Methods — Transformer-based models significantly improve electrical load forecasting accuracy across grid levels.
• Behaviour-Conditioned Neural Processes for Adaptive Residential Short-Term Load Forecasting — Embedding inferred behavioural structure within a Neural Process enhances adaptive residential short-term load forecasting.
• Bet on Features: Anytime-Valid and Feature-Aware Auditing of Conditional Quantile Forecasters — A new framework enables continuous, feature-aware auditing of conditional quantile forecasts, addressing information asymmetry.
• SPECTRA: State-Space Exogenous Context and Temporal-Frequency Resolution Architecture for Probabilistic Energy Forecasting — Probabilistic energy forecasting benefits from explicitly decoupling predictable structures from uncertainty-bearing fluctuations.
• Censoring-Aware In-Context Learning for Generalized Supplier Lead Time Estimation in Supply Chain Planning — LT-ICL uses censoring-aware in-context learning for accurate, low-adaptation-cost probabilistic lead time forecasting with right-censored data.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.