SolarDuel runs a motor-driven tracker and a fixed reference panel side-by-side under identical irradiance, then quantifies the energy difference with a calibrated INA226 power monitor. The goal isn't just to track the sun — it's to prove the gain, with numbers.
- What Makes This Different
- System Architecture
- Bill of Materials
- Hardware Schematic
- System Photo
- Control Pipeline — The Engineering Core
- Signal Flow Diagram
- Logger & Dashboard
- Efficiency Analysis Results
- Repository Structure
- How to Run
- 3D Print Settings
- License
Most LDR trackers on the internet are a four-resistor bridge wired into an analogWrite() loop. SolarDuel treats the same problem as a real control engineering exercise:
| Concern | Naïve tracker | SolarDuel |
|---|---|---|
| Sensor noise | Raw analogRead() |
Per-channel Kalman filter + ADC oversampling |
| Hunting at boundary | Single threshold → chatter | Adaptive deadband with 2.5× hysteresis exit |
| Twitching on light flicker | Acts on every blip | Debounced activation (confirmation counter) |
| Derivative kick on setpoint | None / unfiltered | Derivative-on-measurement with bilinear LPF |
| Integrator runaway at servo limit | Windup → overshoot | Trapezoidal integration + conditional anti-windup |
| Mechanical snap | Step PWM commands | Slew-rate limiter (15 μs / loop) |
| Boot-time shock | Servo snaps to home | 1-second soft-start ramp |
| Loop timing jitter | delay() in loop() |
Fixed 200 Hz scheduler, tick-overflow safe |
| Data acquisition vs. control | Same MCU → coupling | Two isolated MCUs — STM32 controls, ESP32 observes |
The result is a servo that moves like it has inertia — smoothly, decisively, and without the cricket-leg twitching that plagues hobby trackers.
SolarDuel is deliberately split into two electrically and logically isolated units. The tracker never knows it's being measured; the logger never affects motor timing.
┌──────────────────────────────────────────────────────────┐
│ TRACKER UNIT (STM32) │
│ │
│ 4× LDR ──ADC──► STM32-G031K8 ──PWM──► 2× Servo (Pan/Tilt)
│ │ │
│ └─ Kalman → PID → Slew → Soft-LPF │
└──────────────────────────────────────────────────────────┘
┌──────────────────────────────────────────────────────────┐
│ LOGGER UNIT (ESP32) │
│ │
│ Panel ──► 47 Ω load ──► INA226 ──I²C──► ESP-WROOM-32 │
│ │ │
│ ├─► WiFi → Browser dashboard
│ └─► /csv → Timestamped CSV
└──────────────────────────────────────────────────────────┘
| Component | Role / Specification | Qty |
|---|---|---|
| STM32-NUCLEO-G031K8 | Tracker MCU — LDR sampling, control loop, PWM | 1 |
| ESP-WROOM-32 | Logger MCU — INA226 readout, web server, CSV export | 1 |
| LDR (GL5528 class) | Cross-placed light sensors, 3.3 V dividers | 4 |
| INA226 | 16-bit I²C power monitor, 0.1 Ω shunt | 1 |
| Servo (MG996R / SG90) | Pan + tilt actuators | 2 |
| Solar panel | 6 V / 150 mA monocrystalline | 2 |
| Dummy load | 47 Ω 5 W ceramic resistor (MPP-adjacent operating point) | 1 |
| Resistors | 10 kΩ LDR voltage dividers | 4 |
| Power supply | 5 V / 3 A external adapter (servos) | 1 |
| Mechanical | Custom 3D-printed pan-tilt frame, PLA | — |
Full wiring for both units — tracker (STM32 + LDRs + servos) and logger (ESP32 + INA226 + dummy load):
Wiring highlights:
- INA226:
SDA → GPIO21,SCL → GPIO22on the ESP32 - LDRs: 10 kΩ dividers to 3.3 V, outputs to STM32 ADC channels 0, 1, 4, 5
- Servos powered from the external 5 V adapter — never from the STM32 3.3 V rail
- 47 Ω dummy load sits across the panel terminals to pull operation toward the MPP
The control stack is built as a chain of independent stages. Each stage solves one specific real-world failure mode of a basic LDR tracker.
Each LDR is sampled 3 times in software per loop and averaged. Cheap, immediate noise reduction before any filtering.
A 1-D Kalman filter runs on each of the four LDR streams independently:
k.p = k.p + q
kg = k.p / (k.p + r)
k.x = k.x + kg · (z − k.x)
k.p = (1 − kg) · k.p
Tuned with q = 2.0, r = 200.0 — biased toward smoothness because real solar light changes slowly relative to sensor noise.
A fixed deadband fails: too tight under bright sun (chatter), too loose at dawn (poor pointing). SolarDuel scales the deadband with measured ambient light:
tol = TOL_MIN + (ambient / 4096) × (TOL_MAX − TOL_MIN) → tol ∈ [50, 80] ADC units
Brighter ambient → larger differential noise → wider deadband. The controller stays still when it should.
A single noisy sample shouldn't wake the motor. The controller requires CONFIRM_THRESH = 2 consecutive out-of-band samples before activating. Transient cloud edges no longer trigger pointless servo movements.
Not a textbook PID. Four upgrades over the naïve form:
a) Trapezoidal (Tustin) integration
integral += 0.5 · (error + prev_error) · dtMore accurate than rectangular integration at the loop's ~200 Hz rate.
b) Derivative on measurement Differentiating error causes a "derivative kick" on any setpoint change. SolarDuel differentiates the measurement instead — same dynamic response, no kick.
c) Bilinear-LPF on derivative
Raw dD/dt is a noise amplifier. The D-term passes through a first-order discrete LPF:
c1 = (1 − k) / (1 + k), c2 = k / (1 + k), k = α/2
d_filtered = c1 · d_filtered + c2 · (raw_d + prev_raw_d)
Removes the high-frequency chatter that makes textbook PID unusable on servos.
d) Hysteresis exit + asymmetric thresholds Once active, the controller doesn't deactivate at the same threshold it entered:
thr_enter = tol → wake the motor
thr_exit = tol × 2.5 → only sleep when well back inside
This kills limit-cycling at the boundary — a problem most hobby trackers never solve.
PID output is clamped to ±15 μs per loop before being added to the servo target. Mechanically jarring fast moves are physically impossible.
if (Δ > MAX_DELTA_PER_LOOP) Δ = MAX_DELTA_PER_LOOP;
if (Δ < -MAX_DELTA_PER_LOOP) Δ = -MAX_DELTA_PER_LOOP;When the target hits the mechanical limit (PAN_MAX, TILT_MIN, etc.), the integrator is checked: any component pushing further into the limit is zeroed, but useful "pull-back" integral is preserved. Cleaner than blanket clamping.
A subtle but real artifact: casting float → uint16_t truncates, creating 1-step quantization jitter. All PWM writes use:
__HAL_TIM_SET_COMPARE(&htim1, ch, (uint16_t)(pos + 0.5f));The + 0.5f converts truncation into proper rounding. Last source of micro-stepping eliminated.
On boot, servos ramp from PAN_MIN/TILT_MIN to home over 50 steps × 20 ms = 1 second, instead of snapping. Prevents power-up inrush and mechanical shock.
Loop runs at exactly 200 Hz (LOOP_PERIOD_MS = 5) using non-blocking HAL_GetTick() comparison with signed-cast tick-rollover handling (49-day wrap is correctly handled). If execution falls more than 100 ms behind (e.g., debugger pause), the scheduler resyncs instead of catch-up-bursting.
4× LDR
│
▼ ADC oversampling (3 samples / channel)
Kalman filter (per channel, q=2.0, r=200.0)
│
▼ computed each loop
Edge averages → ambient light → adaptive deadband
│
▼
Differential error (right−left) and (bottom−top)
│
▼
Debounce confirm counter (CONFIRM_THRESH = 2)
│
▼
PID — Tustin integration + filtered D-on-measurement
│
▼ ±MAX_DELTA_PER_LOOP
Slew rate limiter
│
▼
Target update → conditional anti-windup at limits
│
▼ (+ 0.5f) cast
Float-rounded PWM compare register write
│
▼
Servo
The ESP32 logger runs an independent HTTP server + live JSON API on the local network. No cloud, no app, no broker — open the IP in any browser.
Features:
- 🟢 1 Hz refresh of voltage, current, power, energy, efficiency
- 📈 Live 60-second rolling power graph rendered on canvas
- ⏺ One-click CSV recording (
Time, V, mA, mW, Wh) with timestamped filename - 🔄 Reset clears energy accumulator and graph buffer
- 📊 Up to 3600 entries (1-hour session at 1 Hz) buffered in RAM
- ⚡ Efficiency is computed against the panel's rated
PANEL_MAX_W = 0.9 W
The dashboard is served as a single in-flash PROGMEM HTML page with vanilla JS — no external libraries, no build step.
A 3-hour A/B session (10:00–13:00) under natural sunlight, panels swapped under matched irradiance, data processed in MATLAB.
| Metric | Tracker | Fixed Panel |
|---|---|---|
| Peak Efficiency | 74.44 % | 35.00 % |
| Average Efficiency | 52.31 % | 21.14 % |
| Peak Power | 670.0 mW | 315.0 mW |
| Average Power | 470.8 mW | 190.2 mW |
| Total Energy Harvested | 1.4125 Wh | 0.5707 Wh |
| Tracker Energy Gain | — | +147.5 % |
| Avg. Instantaneous Gain | — | 180.3 % |
(a) Instantaneous power output. The tracker consistently delivers more power, with both panels showing characteristic cloud-induced dips. (b) Panel efficiency relative to the 0.9 W rated maximum. (c) Cumulative energy harvested. (d) Instantaneous tracker gain over fixed — averaging ~180 % highlights the value of continuous sun-facing orientation.
SolarDuel/
├── CAD_Files/STL/ # Print-ready STL files for pan-tilt structure
├── Hardware/
│ └── hardware_schematic.pdf # Full wiring schematic
├── Docs/
│ ├── dashboard.png # Web dashboard preview
│ └── system_photo.jpeg # Physical system photo
├── Software_STM32_Control/ # STM32CubeIDE project — tracker firmware (C, HAL)
├── Software_ESP32_Logger/ # Arduino IDE project — INA226 + web dashboard
└── Data_Analysis/
├── tracker_panel_data.csv # Logged data — tracking panel
├── fixed_panel_data.csv # Logged data — fixed panel
├── efficiency_analysis.m # MATLAB analysis script
└── efficiency_analysis.png # Output figure
1 — Tracker (STM32)
- Open
Software_STM32_Controlin STM32CubeIDE. - Compile and flash to the NUCLEO-G031K8.
- Power LDRs from 3.3 V, servos from the external 5 V adapter.
- On boot, the soft-start routine ramps the servos to home over ~1 s.
2 — Logger (ESP32)
- Open
Software_ESP32_Loggerin Arduino IDE. - Install the
INA226_WElibrary from Library Manager. - Set your WiFi credentials at the top of the sketch.
- Flash to the ESP32; wire
SDA → GPIO21,SCL → GPIO22. - Connect the active panel to the INA226 input, with the 47 Ω dummy load to ground.
- Open Serial Monitor (115200 baud) — the assigned IP is printed on boot.
- Browse to
http://<IP>from any device on the same network.
3 — Data Collection
- Click ⏺ Start Recording on the dashboard.
- Click ⏹ Stop & Download CSV when done — file is saved with an ISO timestamp.
- Swap panels under matched conditions and repeat for the A/B pair.
4 — Analysis
- Place both CSVs and
efficiency_analysis.min the same folder. - Run the script in MATLAB — summary table prints to console, figure saves as
efficiency_analysis.png.
| Parameter | Value |
|---|---|
| Printer | Creality CR-10 Smart Pro |
| Slicer | Creality Slicer |
| Material | PLA |
| Infill | 15 % |
| Supports | Per-part — check STL orientation |
Attribution: STL files in
CAD_Files/STL/are sourced from this Instructables project and are not original work. All credit goes to the original author. Files are shared here under the terms of the original license for non-commercial, educational use only.
This project is licensed under the MIT License — see the LICENSE file for details.
SolarDuel is an applied engineering study at the intersection of
automatic control theory · embedded systems · data-driven performance analysis.