Can AI forecast a city’s power demand from ten annual observations? Tiny reported errors rest on a two-year test
A peer-reviewed study forecasts electricity demand in Adigrat, Ethiopia, through 2035 with a PSO-optimised neural model. Its total-load error was reported as 0.027%, but the model learned from eight annual observations and was tested on only two.
The short version
- What happened
- The paper uses ten annual observations from 2016–2025: eight for training and two for testing, with domestic, commercial, industrial, street-lighting and total demand modelled separately.
- Who it affects
- Grid planners gain a locally built scenario, but should budget for much larger errors than a two-year back-test can reveal.
- What remains unproven
- Only ten annual observations were available: eight for training and two for testing.
Keep in mind
Only ten annual observations were available: eight for training and two for testing.
There was no external validation on another city, utility or later prospective period.
Interpolation, smoothing and flexible optimisation may make a short trending series easier to fit.
Annual load does not capture operational peaks, feeder constraints, outages or unmet demand.
The 2026–2035 forecast depends on future socioeconomic and climatic inputs that are themselves uncertain.
Source analysis · 1 linked source. Primary analysed source date: 7 November 2026.
Read the full report · 9 min The evidence, methods, implications and limitations
The answer: the model reports exceptional accuracy, but the test is too small to prove general reliability
Researchers modelled annual electricity demand for Adigrat in northern Ethiopia and reported extremely small errors from a hybrid neural system whose parameters and input choices were optimised with particle swarm optimisation. For total load, the paper reports mean absolute percentage error of 0.027%, root mean squared error of about 0.002 and an R² of 0.9999. The same architecture produced reported MAPE values of 0.0172% for domestic demand, 0.094% for commercial, 0.0134% for industrial and 0.052% for street lighting.
Those numbers look unusually precise, but their denominator is the decisive fact. The dataset contains ten annual observations, from 2016 through 2025. The researchers trained on 2016–2023, eight data points, and tested on 2024–2025, two data points. A model can fit a smooth, short annual series closely without demonstrating that it will survive a recession, conflict, a large new industrial connection, tariff reform, extreme weather or measurement changes. Two held-out years do not reveal the likely distribution of future error.
The study is still useful as a worked local planning model in a region often under-represented in energy-AI research. It compares several architectures, documents inputs and publishes decade-ahead estimates. The defensible conclusion is that the PSO-optimised system fit this small Adigrat series very closely. It is not evidence that a utility can expect hundredths-of-a-percent forecasting error in routine operations or in another city.[1]
What went into the forecast
The outcome series came from the Adigrat electricity substation or utility and related administrative records. The paper separates domestic, commercial, industrial, street-lighting and overall electricity demand. Candidate explanatory variables include regional gross domestic product, GDP per person, regional and local population, investment, urban area, humidity, temperature and load factor, alongside time. Missing values were filled by linear interpolation, and a rolling standard deviation was used in handling abnormal metering points.
Five approaches were compared: a conventional regression model, long short-term memory network, LSTM combined with a convolutional network, LSTM combined with a neural network and gated recurrent unit, and an LSTM-neural model optimised by particle swarm optimisation. Particle swarm optimisation searched weights, hyperparameters and features. That makes the winning system flexible relative to the number of observations, increasing the importance of independent validation.
Annual aggregation can be appropriate for long-term capacity planning, but it removes the seasonal peaks that determine whether networks and generation can meet demand hour by hour. A yearly forecast cannot by itself answer how much reserve is needed on a hot evening, which feeder is constrained, or when renewable output and demand diverge. The paper addresses long-range consumption, not operational dispatch, reliability events or household affordability.[1]
The distinction that matters
What the accuracy claim rests on
The reported error is real within the paper's test, but the small denominator sharply limits what it can establish.
Observed history
10 annual points, 2016–2025
A short, smooth series
Model fitting
8 years used for training
Many inputs and flexible architectures
Held-out test
2 years used for evaluation
Total MAPE reported as 0.027%
Unanswered test
No second city or utility
Generalisation remains unknown
Why a two-year test can make performance look more certain than it is
A train-test split is meant to ask whether a model works on observations it did not fit. Here, the test contains two annual outcomes. A percentage error averaged over two points can be tiny when the underlying trajectory is smooth and the explanatory variables track the same trend. It cannot show whether the error is stable, approximately unbiased or robust to a structural break. The near-perfect R² is also less informative on a short trending series than it would be across many independent conditions.
The paper does not provide an external test on a second Ethiopian city or utility. Nor does it use rolling-origin validation across many forecast years, which could at least show how performance changes when the training cut-off moves. Hyperparameter and feature optimisation on such a small dataset can select a configuration that suits this particular history. That risk remains even when the held-out years were not used directly for fitting.
The uncertainty intervals for the 2026–2035 forecast are derived from the model and residual history, but the history itself is only a decade long. Such intervals generally capture variation resembling observed errors; they do not automatically include political disruption, major infrastructure additions or economic pathways outside the sample. A narrow statistical interval should therefore not be read as a complete range of real planning uncertainty.[1]
What the 2035 projection can and cannot tell planners
The study projects load through 2035 for each customer group and in total. These figures are conditional on future values or assumed paths for population, GDP, investment, temperature, humidity and other inputs. If those drivers differ from the scenario, the electricity forecast changes too. The model does not cause future economic growth to occur, and it cannot determine whether suppressed demand, outages or affordability constraints are hiding the amount of electricity people would use if supply were reliable.
For a utility, the most useful role is as one scenario alongside engineering and policy judgement. Planners could compare the hybrid forecast with a transparent baseline, high- and low-growth cases, announced industrial projects and feeder-level constraints. Decisions with long lifetimes should be stress-tested against forecast error far larger than the two-year back-test suggests. A model that predicts annual energy accurately may still miss the location, timing and peak shape that determines investment.
For households and businesses, under-building can mean outages or delayed connections, while over-building can raise costs through underused assets. Transparent assumptions matter because people bear both risks. Publishing input scenarios, revisions and realised errors each year would let communities see why a plan changed and whether the model adds value over simpler methods.[1]
The evidence check
The article is peer reviewed and the authors report no external funding and no competing interests. The comparison includes conventional and deep-learning alternatives, and it supplies customer-category error measures rather than only a headline total. These are strengths. The strongest constraint is statistical: ten annual observations provide very little independent information for a complex hybrid model, especially when multiple predictors, architectures and optimisation choices are considered.
Interpolation and smoothing also deserve attention. Filling missing values and treating anomalous meter readings may improve data usability, but can reduce visible variation and make the series easier to fit. That is not necessarily inappropriate; utilities often need data cleaning. A replicable audit should show which observations were changed, how results move without smoothing, and whether the selected method still wins under alternative cleaning rules.
Reported training time was 557 seconds and inference about 178 milliseconds. That may help reproduce the computation, but speed is not the main deployment bottleneck. Data governance, maintaining future inputs, detecting drift, explaining revisions and integrating forecasts into accountable planning are likely more important than milliseconds of prediction time.[1]
What would change the assessment
Confidence would rise with a longer monthly or quarterly history, pre-specified rolling-origin tests and comparison against simple seasonal, trend and econometric baselines. External validation on other Ethiopian utilities would show whether the architecture learns transferable demand relationships or only Adigrat's short trajectory. A preregistered evaluation should freeze the model before new years arrive and publish every realised forecast error.
Further work should separate energy demand from outages and unmet demand, report peak as well as annual consumption, and test sensitivity to input scenarios. Planners also need cost-weighted errors: underestimating a system peak may matter more than a small annual overestimate. Independent replication with accessible data and code would make the claimed advantage easier to audit.
Until then, the practical reading is cautious. The paper offers a locally grounded demonstration and a plausible way to combine socioeconomic and climatic inputs. Its near-perfect test scores come from two years and should not be used as a promise of precision. The forecast is most valuable as a structured scenario that can be challenged, updated and compared with reality—not as an automatic answer to how much electricity Adigrat will need in 2035.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring electricity demand: compare the finding and its evidence limits.
A two-stage AI cut day-ahead load errors in two datasets—would it help a live grid?
The model produced lower forecast errors than five neural baselines on hourly data from Tamil Nadu and Panama. It was a retrospective benchmark, not a live dispatch, reliability or emissions test.
Read the next reportExplore the full source and evidence record
What this means for people
- Grid planners gain a locally built scenario, but should budget for much larger errors than a two-year back-test can reveal.
- Households and businesses bear the costs of both under-built and over-built infrastructure, making assumptions and annual forecast revisions important public information.
- Researchers in data-scarce regions need methods that communicate uncertainty honestly rather than treating near-perfect fit as proven transferability.
Global context
Long-term demand forecasts are especially important where electrification, urban growth and constrained infrastructure interact. The Ethiopian case broadens geographic coverage, but evidence from one small annual series cannot be transferred automatically to other African grids or countries with different data quality and demand drivers.
What the evidence does not yet show
- Only ten annual observations were available: eight for training and two for testing.
- There was no external validation on another city, utility or later prospective period.
- Interpolation, smoothing and flexible optimisation may make a short trending series easier to fit.
- Annual load does not capture operational peaks, feeder constraints, outages or unmet demand.
- The 2026–2035 forecast depends on future socioeconomic and climatic inputs that are themselves uncertain.
What to watch next
- Prospective performance when 2026 and later demand become available.
- Rolling-origin tests and comparison with transparent statistical baselines.
- External validation on other Ethiopian utilities and higher-frequency data.
- Sensitivity of the 2035 outlook to GDP, population, investment and climate scenarios.
- Public reporting of peak demand, outages and unmet demand alongside annual consumption.
Living evidence record
Impact record IAI-1S7Q03M
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
7 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Scientific Reports published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 7 November 2026
Research question, topics and regions
A small-sample hybrid neural forecast of annual electricity demand in Adigrat, Ethiopia
Topics: Electricity demand · Grid planning · Forecasting · Machine learning · Infrastructure
Regions: Ethiopia · Africa · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.