Watt

React, TypeScript, FastAPI, SQLAlchemy, SQLite, MLflow, XGBoost, Google Cloud Run, Firebase Hosting

Watt is a full-stack web application developed to import, visualize, and forecast home energy consumption: daily import of consumption, index, and peak power data, trend visualization, and future consumption forecasting using machine learning.

This project covers the entire chain: data aggregation from multiple heterogeneous sources (electricity provider, solar panels, electric vehicle, home sensors, weather, water), secure backend API, training and monitoring of predictive models, and automated cloud deployment.

Features

• Dashboard with key indicators (total/average consumption, peak power) and monthly consumption graph
• Detail pages by source: EcoFlow (solar/battery), Enedis (meter index, peak power), Tesla (vehicle charging), Climate (temperature/precipitation), Cost
• Feature Engineering: client-side Pearson correlation matrix and server-side calculated variable importance graph (XGBoost)
• Predictions: backtest comparison between models (Chronos2, TimesFM, XGBoost, Prophet, seasonal average) with MAE/RMSE/bias/MAPE metrics, and production forecast with p10/p90 uncertainty bands
• LLM Analysis: report generated locally by a Qwen3-4B model, comparing backtests and explaining anomalies in French
• Raw SQLite database viewer
• Google Authentication (Firebase Auth) with server-side verified email whitelist

Architecture & Technologies

• Frontend: React, Vite, TypeScript, Tailwind CSS, React Query, Recharts
• Backend: FastAPI, SQLAlchemy 2.0, SQLite
• Machine learning: MLflow (model tracking and registry), XGBoost, TimesFM, Chronos2, Prophet
• Local LLM: Qwen3-4B (GGUF) via llama-cpp-python
• Authentication: Firebase Auth (Google Sign-In)
• Infrastructure: Google Cloud Run, Artifact Registry, Cloud SQL, Firebase Hosting, Terraform
• CI/CD: GitHub Actions (build/test + deployment via Workload Identity Federation, without JSON key)

Key Highlights

• Design of a multi-source data pipeline (Enedis, EcoFlow, Tesla, Aqara, water meter reading, weather) from raw scraping/import to a unique, cleaned, and enriched dataset, ready for training
• Advanced Feature Engineering (solar yield per hour of sunshine, residuals vs trend, thermal amplitude) feeding a correlation matrix and an XGBoost variable importance model
• Training, monitoring, and comparison of multiple time series forecasting models (XGBoost, TimesFM, Chronos2, Prophet) via an MLflow server deployed in production
• Generation of analysis reports in French by a locally executed LLM, integrated as read-only in the application
• Implementation of a complete cloud infrastructure (Cloud Run for the API and MLflow server, Firebase Hosting for the frontend) with automated continuous deployment and static keyless authentication