Behaviour-Conditioned Neural Processes for Adaptive Residential Short-Term Load Forecasting

TL;DR

The paper introduces a behaviour-conditioned Attentive Neural Process model that improves residential short-term load forecasting by inferring household behavioural patterns from context data and using them to condition the forecast decoder. It achieves better accuracy and uncertainty calibration than baseline models, especially with limited context, without requiring labelled behavioural data during training.

Why It Matters

Accurate and robust load forecasting is essential for energy grid management and demand response systems. This approach enables a single probabilistic model to adapt to diverse household behaviours without relying on predefined labels, improving scalability and reliability in real-world deployment.

Key Findings

• The model uses a discrete latent variable to represent inferred behavioural classes, improving forecast accuracy by conditioning the decoder on context-derived behaviour patterns.
• It outperforms both label-agnostic and deterministic baselines in MAE and CRPS across multiple forecast horizons and context lengths, with the largest gains when context is limited.
• The approach maintains lower RMSE than fixed-window models while preserving competitive MAE, indicating fewer large errors under heterogeneous consumption patterns.