Source
https://openenergymaps.org/docs/OEMaps_Documentation_20240326.pdf — original source (opens in a new tab; the file is not redistributed)
Summary
MIT/Power Africa/IEA-developed open data platform (March 2024, V1.0) providing building-level electricity access and demand estimates for low- and middle-income countries via remote sensing + machine learning. Two probabilistic models: BEACON (Bayesian Electricity Access ClassificatiON) infers building electrification status from satellite-derived features fused with utility ground truth; LltLDF (Load Inference through Lightweight Data Fusion) estimates building-level demand under uncertain meter-to-building mapping. Both use multi-input neural networks embedded in Bayesian network architectures with Metropolis-Hastings inference. Validated in East Africa (Rwanda primary), where BEACON achieves 80.7% accuracy (vs HREA 70.9%, GDESSA 48.3%) and LltLDF outperforms eGUIDE state-of-the-art for demand estimation. Open data (geojson + GitHub) and open source. Underlying methodology: Lee S.J. PhD thesis (MIT 2023). Preserved as a knowledge-base entry for a potential future Neobiome research direction in community-scale energy mapping where utility data is sparse.
Key thesis insights
- Multimodal data fusion as the methodological backbone — BEACON and LltLDF fuse five remote-sensing data sources (building footprints from Google + Microsoft; VIIRS nighttime lights; Esri/Microsoft Sentinel-2 land use; high-resolution Esri satellite imagery 50 cm–1 m; Ookla internet speed) with ground-truth utility meter data via a multi-input neural network. The network output feeds a Bayesian probabilistic graphical model for uncertainty quantification. This is the canonical methodology for any future Neobiome research wanting to characterise community-scale energy patterns where utility data is incomplete — potentially relevant for Māori land / papakāinga / iwi-territory mapping where MBIE coverage may be patchy by design or by data-sovereignty negotiation. OT_023
- The probabilistic output is the thesis contribution, not the point estimate — both models output full probability distributions (Beta for access likelihood, Gamma-Poisson compound for demand), not single values. The mean is the “estimate”; the standard deviation is the uncertainty itself as a first-class output. This is the right epistemological framing for any community-scale technology assessment where exact quantification is structurally impossible. NI’s calculation skill could adopt the same pattern — output
(mean, std, distribution_params)rather than just(value). OT_023 - “Inherent lack of information” caveat is honest about epistemic limits — Lee explicitly flags that “it’s likely impossible to actually know precisely what building electricity access status’ are and electricity demand solely based on remote sensing data alone. Distribution lines are often too small to see from the resolutions of satellite imagery.” This is a model of how to handle epistemic uncertainty in community-scale research — bracket the unknown rather than over-claim. Directly applicable to thesis chapters where Neobiome makes quantitative claims about community SS at scales below the resolution of available data. OT_023
- Open-data + open-source philosophy operationalised — entire platform is open: data (geojson), code (BEACON + LltLDF GitHub repos), web interface (openenergymaps.org), and underlying thesis. Sponsored by USAID Power Africa + IEA + Energy for Growth Hub. Citable precedent for a community-data-sovereignty argument in the thesis: when energy-mapping platforms are open, communities can audit, replicate, and challenge the inferences made about them. This pairs naturally with Māori-data-sovereignty frameworks (Te Mana Raraunga) that thesis chapters on indigenous-community SS would engage. OT_023
- Class-imbalance honesty applies to community sampling — Lee flags that East African ground truth has “very few large commercial and industrial consumers that comprise a major share of demand”, creating model bias toward residential. The structural lesson for Neobiome: any community-scale dataset is a sampling problem, and small-N atypical cases (a sawmill, an off-grid Maori arts cooperative, a school) can dominate or distort aggregate statistics. Methodologically applicable across thesis chapters that use community case studies. OT_023
- Potential NZ application — flagged for future return: methodology could be transferred to (a) papakāinga energy demand mapping where utility data has indigenous-territory gaps, (b) off-grid Māori land electrification feasibility, (c) community-scale demand modelling for areas where smart meters don’t yet exist (e.g. rural marae). These are NOT currently active research directions. To activate: (i) confirm whether MBIE/EA data has identifiable Māori-territory coverage gaps; (ii) test methodology transferability outside East Africa via Lee’s GitHub repo; (iii) engage with iwi data-sovereignty principles before any deployment. OT_023
Notes
V1.0 user documentation, March 2024. Living platform — IEA + Power Africa as institutional sponsors; Stephen J. Lee (MIT) as principal author and continuing maintainer (leesj@mit.edu). Web platform updates periodically as new ground-truth data and improved features arrive. All inputs except high-resolution satellite imagery are freely available; satellite imagery procurement is left to the user (Esri commercial). Models can run without satellite imagery (degraded accuracy). Conda environment recommended via Mamba solver.
The methods behind this platform are set out in the underlying Lee (2023) MIT PhD thesis, covered on LIT_052, which carries the per-chapter methodology (Ch5 BEACON access model, Ch6 LItLDF demand model, Ch3 demand-accuracy→cost-of-service sensitivity, Ch2 value-of-probabilistic-forecasts).
Connections
Links to
Sources (1): LIT_052
Referenced by
EDT domains (1): D09: Digital Intelligence & Connectivity
Sources (1): LIT_052