Building-integrated photovoltaics forecasting: Machine learning models for irradiance and power, feasibility, and recommended directions
Building-integrated photovoltaics (BIPV) are central to energy-efficient buildings, yet forecasting irradiance and power with operational practicality remains challenging. This review analyzes 70 studies on machine learning approaches for BIPV forecasting and organizes the literature by task and dep...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | en |
| Published: |
Elsevier
2026
|
| Online Access: | http://eprints.utem.edu.my/id/eprint/29240/2/01941191120252128322525.pdf http://eprints.utem.edu.my/id/eprint/29240/ https://www.sciencedirect.com/science/article/pii/S095219762503146X https://doi.org/10.1016/j.engappai.2025.113115 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Building-integrated photovoltaics (BIPV) are central to energy-efficient buildings, yet forecasting irradiance and power with operational practicality remains challenging. This review analyzes 70 studies on machine learning approaches for BIPV forecasting and organizes the literature by task and deployment needs. For irradiance forecasting, the survey catalogs common contexts (façade/roof, sky/season, urban geometry), input modalities (telemetry, weather, imagery), model families, and frequently reported error metrics. For power forecasting, horizons from intra-hour to seasonal are compared, distinguishing direct power modeling from approaches that incorporate irradiance estimates. Beyond accuracy, the review examines feasibility and robustness, which determine field viability—namely, latency and memory budgets for edge versus cloud execution, sensitivity to outdoor variability, and risks to generalization across sites and seasons. The review concludes with recommended directions focused on implementable next steps, including shared data/metadata conventions for comparability, methods that can be transferred across buildings and seasons, interpretable operations for building-level decision workflows, and pipelines designed for edge execution. The resulting map of models, inputs, and constraints enables readers to match techniques to forecasting tasks and operational contexts. |
|---|