Filtered AI Architecture &
Use-Case Matrix for Solar PV
Mini-Grids in South Asia & SSA

The Engineering Challenge

In isolated Sub-Saharan Africa (SSA) mini-grids utilizing lead-acid or hybrid storage, rapid, unpredicted shifts in cloud cover cause violent, deep battery cycling. If the grid transitions into a Partial State of Charge (PSoC) without warning, irreversible sulfation accelerates battery degradation. Traditional weather frameworks are too coarse (often representing a 100km area) to forecast localized cloud cover, leading to premature battery failures within 18–24 months.

The Open-Source Solution

• Day-Ahead Planning: Quartz Solar Forecast ingests long-range NASA POWER data to generate a baseline generation envelope for the next 24 hours. powerFormer runs simultaneously to project consumer demand. The EMS subtracts the load from the generation profile to pre-allocate minimum state-of-charge (SoC) safety buffers (dynamic SoC limits).
• Real-Time Execution: On-site, a low-cost upward-facing fish-eye camera stream is fed directly into SolarNet. Operating at the edge, SolarNet evaluates cloud velocity and optical thickness to execute intra-hour solar nowcasting.
• Asset Protection: If SolarNet detects a massive localized irradiance drop arriving in 10 minutes, the system stops active battery discharging and dynamically curtails non-essential Productive Use of Energy (PUE) loads (e.g., agricultural water milling) before the battery undergoes a high-current step-discharge.

The Engineering Challenge

In South Asian grid-connected microgrids, phase unbalance is fluid and severe. As evening agricultural pumping spins up on single phases, or as unstructured residential rooftop solar floods individual feeders, the Voltage Unbalance Factor (VUF) easily breaks the 2% safety threshold. This causes localized neutral wire heating, trips commercial inverters, and triggers catastrophic distribution transformer failures due to negative-sequence magnetic strains.

The Open-Source Solution

This architecture establishes an autonomous, decentralized control loop that stabilizes grid voltages without human intervention:

• Step 1 (Topology Mapping): GNN-PowerFlow maps the mini-grid's physical electrical topology into a graph database. Unlike traditional heavy iterative Newton-Raphson solvers, the Graph Neural Network resolves the unbalanced multi-phase power flow equations in milliseconds at the edge.
• Step 2 (Predictive Risk Assessment): Grid AI runs an LSTM pipeline on top of the power flow metrics to identify branches at high risk of voltage sag or phase-neutral thermal overloading within the next hour.
• Step 3 (Cooperative Mitigation): The system invokes PowerNet—which is explicitly architected as an on-policy, cooperative Multi-Agent Reinforcement Learning (MARL) algorithm. Each smart inverter and automated phase-change switch acts as an independent agent. Communicating via PowerNet's differentiable protocol, the agents execute synchronized adjustments:
  • Inverter 1 adjusts its PWM switching pattern to execute asymmetric reactive power ($Q$) injection on Phase A to pull the voltage back up.
  • Smart Static Phase Switches dynamically re-terminate high-load single-phase commercial users to Phase C, instantly flattening fundamental frequency asymmetry.

The Engineering Challenge

Solar mini-grids in rural regions regularly dump up to 30% to 40% of their peak midday generating capacity due to solar curtailment. To stay profitable, operators must actively incentivize "Productive Use" loads (e.g., ice-making, crop drying, EV mini-bus charging) to consume energy exclusively during high-solar windows without overloading the system when solar drops.

The Open-Source Solution

Operators can use GridLearn and DMP-PCFC to transition the grid from a passive generation-following model to an active demand-molding model:

• Load Prediction: DMP-PCFC uses its multi-step time-series architecture to forecast exactly when residential domestic cooking and lighting loads will spike in the evening, defining the grid's "unavailable" energy boundary.
• Demand Response Orchestration: The remaining available midday energy capacity is mapped into GridLearn, an open-source testbed optimized for multi-agent coordination in localized spaces. GridLearn coordinates with smart agricultural loads, treating them as flexible virtual batteries. It schedules water-pumping stations to run sequentially rather than simultaneously, flattening the midday peak and cleanly absorbing excess solar generation without exceeding the thermal limits of distribution transformers.

1. The Free Layer (The Baseline)

The NASA POWER Dataset provides multi-decade global historical databases of solar irradiance, wind speeds, and temperatures. This data is sufficient to train the initial weights of models like powerFormer and Quartz Solar Forecast.

2. The Paid/Third-Party Layer (The Refinement)

Global open datasets carry a coarse spatial resolution (often a 1-degree by 1-degree grid). Mini-grids sit in hyper-local microclimates (e.g., Himalayan valleys in South Asia or coastal fronts in SSA). Operators should subscribe to localized, high-resolution satellite services (such as Solargis or customized Meteostat API streams) to overlay fine-grained local environmental variables onto the global models.