Skip to content

Beyond Weather Correlation: A Comparative Study of Static and Temporal Neural Architectures for Fine-Grained Residential Energy Consumption Forecasting in Melbourne, Australia

Home / Papers / Beyond Weather Correlation: A Comparative Study of Static and Temporal Neural Architectures for Fine-Grained Residential Energy Consumption Forecasting in Melbourne, Australia

Beyond Weather Correlation: A Comparative Study of Static and Temporal Neural Architectures for Fine-Grained Residential Energy Consumption Forecasting in Melbourne, Australia

Authors: Prasad Nimantha Madusanka Ukwatta Hewage, Hao Wu Date: 2026-04-14 Paper ID: openalex:2604.12304

Summary

This study evaluates the importance of temporal autocorrelation versus static meteorological features for fine-grained (5-minute) residential energy forecasting. By comparing MLP and LSTM architectures on two Melbourne households, the authors establish that sequential memory is the dominant predictor for short-term demand. The findings underscore a substantial performance gap between recurrent and static architectures in high-resolution smart meter environments.

Key Contributions

  • Demonstrated that temporal autocorrelation significantly outperforms static meteorological features for 5-minute residential energy forecasting.
  • Quantified performance gaps between LSTM and MLP architectures, showing R^2 improvements of 93.8 and 45.5 percentage points for non-PV and PV-integrated households, respectively.
  • Revealed that static weather-driven models implicitly capture solar generation patterns in PV-integrated dwellings.

Open Questions & Future Work

Archivist Review

The paper provides a straightforward comparative study that lacks novel, reusable architectural concepts or methodologies distinct enough to warrant a standalone vault concept note. However, the identified future research directions regarding solar signal disentanglement and privacy-preserving federated learning represent substantial, persistent challenges in residential time-series forecasting and were therefore elevated to open questions.

Approved Open Questions

  • Solar PV Load Disaggregation: Crucial for addressing the performance bias observed when models implicitly learn solar patterns rather than true energy demand.
  • Federated Learning for Residential Forecasting: Addresses the scalability-privacy trade-off inherent in residential smart grid management.

Rejected Candidates

  • [open_question] Solar PV load disaggregation forecasting (solar-disaggregation-forecasting) - other: This is a re-submission/rewrite of the candidate provided in the analysis; it is approved but required adjustment for the vault format.
  • [open_question] Federated learning for load forecasting (federated-learning-residential-load) - other: This is a re-submission/rewrite of the candidate provided in the analysis; it is approved but required adjustment for the vault format.

Metadata & Links

url
https://arxiv.org/abs/2604.12304
paper_id
2604.12304
paper_source
openalex
domain
time-series
tags
architectures
datasets
concept_slugs
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-04-17T05:46:19Z