An End-to-end Building Load Forecasting Framework with Patch-based Information Fusion Network and Error-weighted Adaptive Loss
An End-to-end Building Load Forecasting Framework with Patch-based Information Fusion Network and Error-weighted Adaptive Loss
Authors: Hang Fan, Ying Lu, Weican Liu, Dunnan Liu, Xiaotao Chen, Shengwei Mei Date: 2026-04-15 Paper ID: openalex:2604.13714
Summary
This paper introduces an end-to-end building load forecasting framework designed to address complex temporal dependencies and high load volatility. The architecture utilizes a Patch-based Information Fusion Network (PIF-Net) to process input series as local patches, employing gated hidden state integration for dynamic temporal weighting. To improve model robustness, the framework is trained with an Error-weighted Adaptive Loss (EWAL) function that adjusts penalties according to prediction error distributions. Experimental results confirm the frameworkâs superior performance in capturing building load patterns compared to baseline models.
Key Contributions
- Proposed the Patch-based Information Fusion Network (PIF-Net) to effectively capture local and global temporal dependencies in load data.
- Developed the Error-weighted Adaptive Loss (EWAL) to enhance prediction robustness against volatile load spikes and extreme conditions.
- Introduced a two-stage preprocessing pipeline incorporating Local Outlier Factor (LOF) and SVM-SHAP for effective anomaly detection and feature selection.
Open Questions & Future Work
Key Concepts
- patch-based-information-fusion-network: A neural architecture for time series forecasting that processes input sequences as local patches and fuses them via a gating mechanism to weight temporal importance.
- error-weighted-adaptive-loss: A training objective for forecasting models that dynamically adjusts penalty weights based on real-time prediction error distributions.
Archivist Review
I approved the two primary architectural and objective function contributions (PIF-Net and EWAL) as they represent reusable paradigms for time-series modeling. I also approved the two open questions concerning the transition to probabilistic frameworks and edge-deployment efficiency, as these represent significant, long-term challenges in the field of time-series forecasting. No new datasets were proposed, and standard preprocessing steps (LOF, SHAP) were rejected as routine practice.
Approved Concepts
- Patch-based Information Fusion Network: It is the core architectural contribution of the paper, enabling multi-scale temporal feature extraction through patching and gated fusion.
- Error-weighted Adaptive Loss: It provides a novel method for improving model robustness in the presence of extreme outliers or high-volatility data by dynamically adjusting penalty weights.
Approved Open Questions
- Probabilistic Building Load Forecasting: Probabilistic forecasting is crucial for robust risk management and grid balancing in the presence of high load volatility, which deterministic models struggle to capture adequately.
- Edge Optimization for Forecasting Models: Edge-based forecasting is necessary for local demand-side management and reducing data transmission latencies, but current deep learning models are often too computationally demanding for these environments.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.13714
- paper_id
- 2604.13714
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-forecastingenergy-forecastinganomaly-detectionfeature-selection
- architectures
-
- datasets
-
- concept_slugs
- patch-based-information-fusion-networkerror-weighted-adaptive-loss
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-18T05:34:50Z