All projects

Case study · A* routing for electric vehicles, with live charging stops

EV Route Optimizer

An Android app that plans multi-stop EV road trips end-to-end — real-time charging station insertion via OpenChargeMap, GraphHopper road routing, Open-Meteo weather effects, OBD-II Bluetooth SoC reading, and turn-by-turn MapLibre navigation. Built with Jetpack Compose and a custom A* engine over a (waypoint × SoC) state space.

Kotlin 2.2 Jetpack Compose Material 3 MapLibre Room Koin Retrofit WorkManager
EV Route Optimizer

01

The problem

Every popular EV navigation app treats charging stops as an afterthought — they slot in a single stop along the fastest road, ignoring your current battery level, your charger type, real-time station availability, ambient temperature effects on range, and the compound cost of going above 80% SoC on a tapered charge curve. The result is routes that look fast on paper but leave you range-anxious on the motorway.

02

The approach

Built a custom A* search engine over a two-dimensional state space: (waypoint index × state-of-charge). Each node is a (location, SoC) pair; edges are road segments whose energy cost comes from a physics model accounting for speed-dependent aerodynamic drag, rolling resistance, elevation gradient, temperature derating, and HVAC overhead. The charging model applies a non-linear taper above 80% SoC. Four optimisation modes (Fastest, Fewest Stops, Cheapest Charging, Battery Friendly) weight the heuristic differently. The engine dominance-prunes states, so it never expands a node if a cheaper path to the same location at higher SoC already exists.

03

My contribution

Sole developer — A* engine, energy and charging models, all four API integrations (GraphHopper, OpenChargeMap, Open-Meteo, Photon), OBD-II Bluetooth SoC reader, Room caching, WorkManager prefetch, calendar integration, and the full Compose UI with MapLibre navigation.

04

What I learned

The routing problem is harder than it looks because SoC is not just a constraint — it is part of the search state. Treating it that way (rather than post-processing a road route) is what allows the engine to find routes that would never emerge from a road-distance-only search: a slightly longer road segment that bypasses a detour to a slow charger, or a route that arrives at a fast charger with exactly enough margin to charge quickly and press on. The other hard lesson was the energy model: an overly aggressive speed-consumption curve caused the engine to over-insert stops on motorways. Calibrating it against real-world EV consumption data (80% rolling resistance, 20% aerodynamic drag growing with v²) brought the stop count in line with what drivers actually experience.


Screenshots

Plan Route — origin, destination, battery parameters and connector type Route Itinerary — charging stops, durations, arrival SoC, and quality score Turn-by-turn Navigation — MapLibre map with route line and speed/ETA HUD

A multi-stop EV route planner that treats charging as a first-class citizen of the search — not a post-processing step bolted onto a road-distance result.

EV Route Optimizer plans electric-vehicle road trips end-to-end: you enter origin, destination, battery parameters, and connector type; the engine finds the optimal sequence of charging stops and returns a full itinerary with charge durations, arrival SoC at each stop, and a quality score versus the unconstrained baseline.

The app integrates four live APIs — GraphHopper for road segments and elevation, OpenChargeMap for real-time charging station data, Open-Meteo for temperature and weather, and Photon for geocoding — and reads current state-of-charge directly from the vehicle via an OBD-II ELM327 Bluetooth adapter.

Architecture

Architecture diagram: Compose screens and a MapLibre map sit on a shared PlanViewModel, which uses the RoutingEngine (A*), EnergyModel, ChargingModel and CalibrationStore plus Route, Station, Weather, Geocoding and OBD repositories; those call the GraphHopper, OpenChargeMap, Open-Meteo and Photon APIs, cache in a Room database and read SoC from an ELM327 Bluetooth adapter; a WorkManager worker pre-warms the caches.

How the routing engine works

The A* search operates over a (waypoint, SoC) state space. Every node is a (location, battery%) pair. Edges are road segments whose cost is computed from a physics-based energy model:

  • Speed-dependent drag — rolling resistance (~80%) plus aerodynamic drag (~20%) growing with v². Calibrated to real EV consumption curves so the engine doesn’t over-insert stops on motorways.
  • Elevation factor — net gravitational work per segment, signed (uphill costs, downhill recovers).
  • Temperature derating — cold-battery and HVAC overhead extracted from Open-Meteo forecasts at the departure time.
  • Charging model — non-linear taper above 80% SoC reflects real fast-charger behaviour; cold-battery derate applies to charging speed.

Dominance pruning ensures the engine never re-expands a (location, SoC) pair if it has already found a path to the same location at equal or higher SoC for equal or lower cost.

Four optimisation modes

Mode What it optimises
Fastest Minimise total trip time including charge sessions
Fewest Stops Minimise number of charging stops
Cheapest Charging Minimise charging cost using per-network tariff inference
Battery Friendly Gentler charge cycles and moderate speeds

Results

Metric Value Source in repo
Unit tests 25 (9 charging model, 8 energy model, 8 routing engine) app/src/test/.../domain/engine/{ChargingModelTest,EnergyModelTest,RoutingEngineTest}.kt
Optimisation modes 4 domain/model/Models.kt (OptimizationMode)
SoC resolution in the search 5% buckets RoutingEngine.kt (socStepPercent)
Documented routing scale Dominance pruning keeps the search tractable on 250+ km routes README.md
Live data sources 4 APIs (GraphHopper, OpenChargeMap, Open-Meteo, Photon) data/remote/api/, README.md
Route and station cache TTL 6 hours RouteRepository.kt, StationRepository.kt

The repository documents no benchmark timings or dataset sizes, so none are claimed here.

Key features

  • Live charging station insertion — OpenChargeMap filtered by connector type, minimum power, and a 6-hour freshness TTL.
  • OBD-II SoC reading — ELM327 Bluetooth adapter; non-blocking I/O with a 3.5s timeout; circuit-breaker after 5 consecutive failures.
  • Offline fallback — Room-cached routes (55m coordinate tolerance) and stations serve results when the network is unavailable.
  • WorkManager prefetch — background job pre-warms station and route caches for saved trips before departure.
  • Calendar integration — reads upcoming events with location fields and surfaces them as one-tap trip suggestions.
  • Turn-by-turn navigation — MapLibre map with route polyline, location tracking, and a speed/ETA heads-up display.
  • Sensor calibration — persistent efficiency calibration from real trips to refine the energy model over time.

Engineering highlights

  • Custom A* over a two-dimensional state space — not a wrapper around any existing routing library.
  • Physics energy model calibrated to real EV consumption data, not a lookup table.
  • Non-blocking OBD serial I/O (available() polling + deadline loop) to avoid indefinite blocking on InputStream.read().
  • Room coordinate matching at 0.0005° (~55 m) to avoid serving cached routes for wrong locations.
  • Dispatchers.Default for CPU-bound JSON serialisation; Dispatchers.IO for all network and database work.
  • MVVM + Repository + Koin DI throughout; WorkManager try-finally ensures DB handles are always closed.

↑↓ to navigate ↵ to open