Source
https://dspace.mit.edu/handle/1721.1/152664 — original source (opens in a new tab; the file is not redistributed)
Summary
Stephen J. Lee’s 2023 MIT PhD thesis develops a suite of probabilistic machine-learning methods for estimating electricity access and demand in low- and middle-income countries, where utility metered data is sparse, noisy, or absent — the data gap that blocks right-sizing the infrastructure needed for UN SDG7. The thread running through it is that, because energy investments are large, hard to reverse, and made under deep uncertainty, probabilistic estimates that quantify their own uncertainty are strictly more valuable than point estimates for planning. The thesis builds this argument across five technical chapters: Ch2 establishes the decision-theoretic value of probabilistic forecasts (real options, value of information); Ch3 uses a REM technoeconomic model in Uganda to show demand-estimation accuracy can swing electricity unit cost ~3× and that coordinating clean-cooking with electrification cuts costs ~34%; Ch4 (AMPED) forecasts country-level annual demand for 43 African countries; Ch5 (BEACON) estimates building-level electrification status in Rwanda at 80.7% accuracy (beating HREA, GDESSA, naive baselines) by fusing satellite imagery, nighttime lights, building footprints, land use and internet-speed data through a neural-network-in-Bayesian-model (LDF) architecture; and Ch6 (LItLDF) estimates building-level demand under noisy many-to-many meter↔building geolocation by embedding the NN inside a Bayes net with MCMC inference. This thesis is the full methodology behind the Open Energy Maps™ platform recorded as OT_023, and resolves RT_116.
Key claims
See key_claims frontmatter (8 claims, per-chapter, all cited to source location).
Neobiome Intelligence relevance
- Probabilistic output as a first-class NI deliverable. The thesis’s central methodological stance — output a distribution (mean + uncertainty), never a bare point estimate, because uncertainty is itself the input planners need to manage risk — is a direct design pattern for the NI calculation layer: emit
(mean, std, distribution_params)rather than a single value. This reinforces the same idea already flagged in OT_023 and pairs with the resource-flow-accounting output discussion on the Self-Sufficiency Calculation Framework. LIT_052 - Demand-estimation accuracy drives infrastructure right-sizing (energy planning, D01/D09). Ch3’s finding that modelled cost of service swings from US
0.13 to0.37/kWh — nearly 3× — purely on the demand assumption, and that modelling consumer heterogeneity yields ~9% cheaper plans, is a quantified argument for why NI must size against a defensible demand profile, not a single nominal load. Method, not NZ cost data. LIT_052 - Value-of-information / real-options framing for staged deployment. Ch2’s result that probabilistic forecasts favour smaller, modular, flexible investments (and the Palawan ~$150M option-value case) supports an NI design where a community can stage technology deployment and preserve optionality under uncertainty — echoing the small-scale-vs-centralised tension in OT_053 and the individual-PV-vs-microgrid siting question in LIT_048. LIT_052
- Coordinated multi-resource planning beats siloed planning. Ch3 Case 3 (coordinating electrification with clean-cooking cuts costs ~34% and lifts electric-cookstove viability 42%→82%) is empirical support for NI treating energy + cooking/heat demand as a coupled system rather than independent domains. LIT_052
- Remote-sensing demand/access inference is a flagged FUTURE NI layer, not an active feed. The BEACON/LItLDF methods would only become NI-relevant in NZ contexts where utility data is structurally unavailable (off-grid Māori land, papakāinga, pre-smart-meter rural areas). The Tasman pilot has higher-quality grid data already (MBIE/CR_009, Network Tasman, NIWA), so these methods are parked as a future direction — consistent with how OT_023 was routed. LIT_052
- ⚠ Scope. African/LMIC empirical study (Rwanda primary; Kenya, Uganda, Mozambique, Palawan secondary). Every numeric result is a method demonstration in that context — none are NZ benchmarks. Building-level accuracy is itself unverified (validated only at aggregated subgraph level), and the authors flag a price-inelastic / reliability-independent demand assumption known to be false.
Key thesis insights
- Epistemic humility about sub-resolution claims. Lee is explicit that building-level electricity status cannot be known precisely from remote sensing (distribution lines are too small to see), so the work brackets the unknown with distributions rather than over-claiming — a model for how thesis chapters should handle quantitative claims at scales below available data resolution. LIT_052
- Open-data / community-data-sovereignty precedent. The entire platform (data, BEACON + LItLDF code, web interface, this thesis) is open — a citable precedent for the thesis argument that when energy-mapping platforms are open, communities can audit, replicate and challenge the inferences made about them (pairs with Te Mana Raraunga / Māori data-sovereignty framing). LIT_052 OT_023
- The regulatory-economics lineage. Supervised by Pérez-Arriaga (a ‘utility approach to accelerate universal electricity access’ and the Global Commission to End Energy Poverty), the thesis sits in an institutional-design tradition that links technical demand estimation to regulatory and tariff policy — relevant background for thesis chapters on governance of community energy. LIT_052
Research targets
Research gaps
- Whether the NI calculation layer formally adopts a probabilistic output (mean + uncertainty + distribution params) rather than point estimates — a tool-design decision tracked with the composite-design RTs (RT_006, RT_016, RT_029). No external document to retrieve.
- Transferability test: whether the BEACON/LItLDF GitHub methods run on NZ inputs for any off-grid / Māori-territory / pre-smart-meter context — remains the future-direction activation trigger raised in OT_023; not an active research direction.
Connections
Links to
Referenced by
EDT domains (1): D09: Digital Intelligence & Connectivity
Sources (1): OT_023