AI-Driven Solar Inverter Failure Prediction & Intelligence Platform

Aubergine @ HACKaMINeD2026

Architecture Overview

This repository contains a full end-to-end Machine Learning and Generative AI pipeline designed to ingest solar inverter telemetry and output actionable, natural-language maintenance risk protocols.

graph TD;
    A[Telemetry CSVs] -->|preprocess.py| B[Data Cleansing & Timeseries Features]
    B -->|target_7d_failure| C(XGBoost ML Models)
    C -->|Feature Importance| D[SHAP Explainer]
    C -->|Risk Score| E{FastAPI Backend}
    D -->|Top 5 Features| E
    E -->|Narrative Request| F[GenAI Intelligence Layer]
    F -->|Plain-English Action Plan| G(Streamlit Dashboard)

Loading

Solution Components

1. Data Preprocessing (src/data/preprocess.py)

• We processed ~600,000 raw telemetry logs covering a 2-year window for 3 critical inverters.
• We handled completely null attributes, interpolated sensor drop-outs, and built a custom target_7d_failure target logic by monitoring subsequent 7-day windows for anomalous alarm_code triggers or extreme drops in expected active power generation.

2. Predictive ML Models (src/model/train.py)

• We trained high-performance XGBoost Decision Trees utilizing scale_pos_weight to combat the intense class imbalanced nature of temporal anomalies.
• We integrate SHAP (SHapley Additive exPlanations) to extract the exact components (e.g. low PV voltage input vs high temp) driving the risk.

3. Generative Insight Engine (src/genai/narrative.py)

• The FastAPI suite passes the Model's Risk Score and the SHAP Feature Array to a generative text handler mapping numerical anomalies to explicit, plain-English mechanical actions for grid operators.

4. Application Suite

• API (api/main.py): Secured via FastAPI & Pydantic. Validated using Pytest unit tests.
• Frontend (dashboard/app.py): Streamlit graphical interface rendering real-time probability indices and the GenAI recommendation strings.

Installation & Execution

# 1. Install Dependencies
pip install -r requirements.txt

# 2. Run Local Environment
uvicorn api.main:app --reload

# 3. Launch UI (New Terminal)
streamlit run dashboard/app.py

Docker

docker build -t aubergine_inverter_ai .
docker run -p 8000:8000 -p 8501:8501 aubergine_inverter_ai

Design Decisions & Future Improvements

• Decision: Used XGBoost over LSTMs to guarantee explicit tabular explainability via SHAP.
• Limitation: Due to hackathon constraints, the GenAI Engine utilizes deterministic mock mapping rather than calling a live GPT-4 Omni API Endpoint.
• Future Improvements: We plan to mount a live Vector DB / RAG system to let operators natively query the system ("Which inverters failed in block A during June?").