- The paper demonstrates that temporal LSTM models significantly outperform static MLPs in sub-hourly energy forecasting with up to 88.3% variance explained.
- It reveals that historical consumption data better captures the stochastic nature of intra-day loads compared to static daily weather features.
- The study highlights the need for disaggregating solar generation in PV households and suggests using sequence models for smart grid applications.
Comparative Analysis of Static and Temporal Neural Architectures for Fine-Grained Residential Energy Forecasting
Problem Space and Motivation
The paper "Beyond Weather Correlation: A Comparative Study of Static and Temporal Neural Architectures for Fine-Grained Residential Energy Consumption Forecasting in Melbourne, Australia" (2604.12304) addresses the critical challenge of achieving sub-hourly accuracy in residential short-term load forecasting (STLF), an essential requirement for smart grid reliability and demand response under Australia's five-minute settlement regime. The distinctive stochasticity of individual household loads, influenced by occupant behavior and distributed energy resources (notably rooftop PV), complicates forecasting at this granularity.
Although weather variables are widely used as exogenous features, the superiority of temporal history over static meteorological descriptors has not been conclusively established for Australian residential grids at 5-minute intervals. This study systematically compares (i) a static Multilayer Perceptron (MLP) trained on daily weather observations and time-of-day features, with (ii) a Long Short-Term Memory (LSTM) model operating exclusively on sliding windows of prior consumption.
Experimental Design and Data Pipeline
The experiment utilizes 14 months (March 2023–April 2024) of 5-minute interval smart meter data for two Melbourne households—one grid-only, one PV-integrated—resulting in approximately 117,000–119,000 samples per dwelling. Weather data from the Australian Bureau of Meteorology includes six daily variables (after exclusion of features with high missingness, notably solar irradiance). All features are subjected to Min-Max normalization, missing values interpolated, and the data divided chronologically for training (80%) and testing (20%).
Feature engineering for MLP comprises daily maximum/3pm/9am temperatures, rainfall, humidity, and time-of-day. The LSTM ingests only sequences of 24 past consumption values (2-hour history), excluding weather covariates for signal isolation. Model architectures are relatively compact, consistent with the fine-grained nature of the data and the limited sample size.
Model Results and Validation
The empirical results decisively demonstrate the dominance of temporal autocorrelation over weather features for 5-minute STLF:
- MLP (Weather + Time) on Grid-Only Household: R2 = -0.055, worse than mean baseline. RMSE = 971.14 W, substantially above daily consumption variability, confirming the inadequacy of daily weather for sub-hourly prediction.
- LSTM (2-hour sequence) on Grid-Only Household: R2 = 0.883—explains 88.3% of variance. MAE ≈ 130–160 W, indicating high fidelity.
- Naïve Persistence Baseline: R2 = 0.878 for House 3, affirming strong short-range autocorrelation.
- MLP on Solar-PV Household (Implicit Solar Forecasting): R2 = 0.410, reflecting improved predictability due to the deterministic diurnal pattern in solar generation.
- LSTM on Solar-PV Household: R2 = 0.865, maintaining superior performance over MLP by 45.5 percentage points.
Performance stratified by season shows minimal degradation for the LSTM, with the strongest diurnal peaks consistently captured. Diurnal and seasonal analysis indicates that intra-day stochasticity remains refractory to weather-driven MLPs; only temporal modeling captures high-resolution structure.
Mechanistic Interpretation and Feature Analysis
The negative R2 for weather-driven MLP on grid-only households exposes a fundamental granularity mismatch: daily weather features offer negligible discrimination across the 288 intra-day intervals. Consumption variability at this granularity is dominated by stochastic events (appliance usage, behavioral patterns), rendering weather variables effectively non-informative except during temperature extremes.
Conversely, for the PV-integrated household, the MLP's improved R2 reflects its ability to forecast the deterministic solar generation component, which correlates strongly with weather and time-of-day features. This conflation underscores the need to separately model grid draw and solar generation to avoid methodological artefacts and support meaningful load forecasting.
Feature correlation analysis confirms weak weather-load coupling for grid-only dwellings but stronger solar-related correlations for total consumption in PV households. Direct solar irradiance remains absent, limiting PV modeling fidelity and motivating future integration of satellite-derived solar datasets.
Practical and Theoretical Implications
The results have direct implications for the design and operation of smart home energy management systems (HEMS) and virtual power plants:
- Model Selection for HEMS: LSTM architectures trained on historical consumption sequences provide significantly stronger predictive accuracy for 5–30 min horizons, enabling fine-grained control for battery dispatch, appliance optimization, and demand response.
- Weather-Driven Forecasting: Daily weather features alone are unsuitable for single-household STLF at sub-hourly resolution. Intra-day or hourly meteorological data (e.g., NWP output, SILO database) would be necessary to enhance forecasting models.
- Solar Forecasting Methods: Joint modeling and disaggregation of load and solar generation is necessary to avoid overfitting weather-driven MLPs to total consumption.
- Federated Learning and Generalization: Extension to multi-household datasets and federated learning approaches (e.g., FedAvg) are imperative to address privacy constraints and obtain population-level generalization.
Limitations and Future Research Directions
Key limitations include the restricted weather feature granularity (absence of hourly temperature and irradiance), lack of a weather-augmented LSTM, single-run model results without statistical significance estimation, and evaluation across only two households. The outlined research roadmap proposes:
- Incorporating weather covariates into multivariate LSTM and Transformer architectures (Informer, PatchTST, TFT).
- Seasonal stratification analysis at the event level (e.g., heatwaves).
- Large-scale validation across heterogeneous residential populations.
- Federated learning deployments for privacy-preserving aggregation.
- Solar disaggregation architectures for explicit modeling.
Conclusion
This comparative study establishes—at 5-minute resolution for Australian households—that temporal sequence modeling vastly supersedes weather-driven static features in forecasting short-term residential electricity demand. The LSTM architecture achieves up to 93.3 percentage points higher R2 than weather-only MLP, illustrating that stochastic intra-day variability is dominant at this granularity. Implicit solar forecasting is a methodological artefact for PV dwellings, necessitating disaggregation in future work. The clear practical implication is that sequence models should be deployed for sub-hourly residential energy forecasting in smart grids, with weather augmentation reserved for exceptional temperature events. The methodological rigor and published benchmarks establish a foundation for future research integrating advanced time-series models, high-resolution meteorological data, and federated approaches in residential energy analytics (2604.12304).