GPS-only energy prediction for a commercial electric motorcycle fleet using an LSTM neural network

Commercial electric two-wheeler fleets are expanding rapidly across Sub-Saharan Africa, yet the basic operational question of how much energy a given trip will consume remains difficult to answer from the data that commercial fleet operators actually collect. Boda-boda operators run on thin margins, earning on the order of USD 10 per working day with fuel consuming close to half of gross revenue, so the per-kilometre running cost is the decisive quantity in any electrification decision; physical pilot programmes that could quantify that cost do not scale to the cities where the decision is being taken. Onboard battery telemetry is proprietary and unavailable for the internal-combustion vehicles against which electric fleets must be benchmarked, and commercial ICE motorcycles in the region carry no telematics beyond a GPS tracker fitted for anti-theft and fleet management. Standard GPS tracking, by contrast, is already deployed on virtually every commercial motorcycle in the region, which makes it the only signal common to both vehicle types. This paper develops and validates a GPS-only energy prediction methodology for a commercial electric motorcycle (Roam Air) operating as part of a delivery and motorcycle-taxi fleet in Nairobi, Kenya. Twenty-three input features derived entirely from the GPS trace, from SRTM elevation data, and from OpenStreetMap road classifications are fed to a two-layer long short-term memory (LSTM) network trained on per-timestep efficiency (Wh/km). A noise-aware P90 target clipping strategy suppresses battery management system (BMS) tail noise without discarding high-consumption trip segments. On a test set of 7245 trips from 25 riders whose trips were never seen during training (a held-out-rider split), and after a single global calibration constant fitted on the validation set, the model achieves a trip-level coefficient of determination of R 2 = 0.949 ± 0.003 and a mean absolute error of 22.0 ± 0.1 Wh (mean ± std across three independent random seeds, all metrics timestep-weighted), reducing trip-level MAE by 30% relative to a constant Wh/km baseline calibrated on the same validation set. The fleet-level aggregate energy error is −2.2%, and the GPS-only LSTM resolves a Wh/km variance share ( R Wh / km 2 = 0.34 ) that a fleet-average predictor cannot resolve at all ( R Wh / km 2 ≈ 0 ). Against a benchmark suite of linear, random-forest, tuned-XGBoost, and GRU models, the sequence models hold the best trip-energy variance resolution while only they provide the per-step spatial predictions that corridor-level and swap-network planning require Because the method works from data commercial fleets already collect, on freely available public layers, it acts as a virtual pilot for total-cost-of-ownership analysis, battery-swap network sizing, and EV–ICE comparison in emerging markets where the corresponding instrumented studies do not scale.