Optimizing Solar Trackers: On-Chip ML & Bare-Metal STM32

Renewable energy systems are highly dynamic. Solar photovoltaic (PV) arrays generate energy based on solar irradiance and cell temperature. However, under rapid shadow transitions (such as passing clouds or structural shading), the power-voltage curve exhibits multiple peaks.

Traditional Maximum Power Point Tracking (MPPT) models, like standard *Incremental Conductance* or *Perturb & Observe*, routinely get locked into local peaks, failing to capture the global maximum and losing up to 30% of energy capacity.

During the global embedded design challenge at IIT Bombay International Techfest (where we placed 7th Internationally), we designed a solution: Integrating ML directly on bare-metal controllers to dynamically map solar arrays.

Hardware Architecture

Our system is composed of an Interleaved Boost Converter controlled by an STM32 MCU. The interleaved topology was chosen because:

• **Reduced Ripple**: Splitting the input current between two parallel converter phases reduces current ripple, minimizing input capacitor strain.
• **High Efficiency**: Allows components to operate at optimal thermal thresholds under high-wattage inputs.
• **High Reliability**: Offers redundant paths in case of phase component failure.

Deploying Machine Learning on Microcontrollers

Running machine learning models on a resource-constrained STM32 microcontroller is a massive math exercise. General microcontrollers lack dedicated Floating Point Units (FPUs) or heavy vector processing units.

To solve this, we:

1. Mathematical Optimization: Formulated a Multilinear Regression model utilizing standard integer-only matrix algebra. We bypassed heavy floating-point imports by scaling parameters by a factor of 1000 and running arithmetic calculations using fast, bit-shifted integer operations.

2. Dynamic Search Spaces: The on-chip ML model constantly evaluates temperature and ambient light. Instead of conducting exhaustive sweeps across the full voltage range, it instantly predicts the precise coordinate of the global maximum power point and shifts the duty cycle.

3. Hybrid Tracking: Once the ML model positions the converter near the predicted maximum, the system switches to highly precise, micro-conductance sweeps to lock the peak exactly.

Measurable Results

In tests, our hybrid ML-boost prototype delivered:

• **15% Increase** in solar power capture during dynamic, rapid cloud shadow simulations.
• **40% Reduction** in tracking lock times compared to standard Incremental Conductance sweep models.

Conclusion

This project demonstrated that advanced computing is not restricted to high-end cloud servers. Deployed on-chip and optimized at the bare-metal level, compact machine learning models can deliver substantial enhancements to energy conversion efficiency, scaling green infrastructure at low commercial costs.