Project Design & Implementation Document

Demo

Grid.Demo.mp4

1. Project Background & Motivation

With recent events causing Jamaica's power grid to fail for long periods, particularly caused by the worsening weather patterns. My aim was to develop a full-stack, intelligent power grid monitoring and auxiliary decision-making platform. By integrating a power grid simulation framework with machine learning-based anomaly detection and load prediction. The system enables early warning of potential grid failures (line overloads, voltage violations) and generates standardized anomaly analysis reports.

2. System Architecture Design

• Frontend: A Single Page Application built with Vite + React + TypeScript, responsible for spatial topology visualization and time-series data rendering.
• API: An asynchronous RESTful API built with FastAPI. FastAPI was chosen for its asynchronous nature which performs well for high-frequency time-series data streams.
• Simulation & Algorithm: The core physics engine utilizes Grid2Opcombined with a two machine learning models, predicting the value of the highest load in the next 30 mins (~6 steps in Grid2op) and anomaly detection.
• Database: PostgreSQL + TimeScaleDB for historical time-series data persistence; Open-Meteo API for real-world meteorological boundary conditions.

3. Tech Stack

Component	Tech Stack
Simulation	Grid2Op (l2rpn_neurips_2020_track1_small)
Backend	FastAPI + Uvicorn
Frontend	React + TypeScript + Leaflet
Database	PostgreSQL + TimeScaleDB
Reports	WeasyPrint + Jinja2

4. Features

• Grid Simulation : Using Grid2Op to simulate a realistic power grid.

• Real-Time Grid Topology : Using the user's location and their countries substation data gathered from earth-osm to recreate the country's ten closest substations to the user to generate a grid topology on react-leaflet, as well as creating a ring-based grid system if the country's data is not found.

• Anomaly Detection & Highest Load Prediction : Based on features including highest load, the power production for both renewable and non-renewable generators, load drop, rolling mean and standard deviation, the difference between the power generated vs the power used and the active/reactive power consumption. These features were generated from Grid2Op to train a LGBMClassifier to predict anomalies within the grid and a LGBMRegressor to predict the highest load in the system in the next 30 mins (~6 steps in Grid2Op).

• Top 5 Load Time Series Chart : A time series plot of the top five highest loads in the grid using recharts.

• Report Generation : Generates a report based on the most recent anomalies.

5. Model Evaluations

Grid Anomaly Classfication Model

The dataset held high imbalance, as such a precision-recall curve was the best method to visulalize the performance of the model. The model had strong performance having an average precision of 0.97.

Load Prediction Model

The model held strong performance in the range of 0.65 rho to 0.90 rho, and began underestimating at >0.90 due to the rarity of these cases from the simulated data. The model made predominant usage of max_rho for predicting the target_rho, but also made use of power production and consumption. The rolling variables had much smaller influence on the prediction. There was only a 1.14% error when predicting, and only ~3.3% of the variation remains unexplained.

Colab used to evaluate model

https://colab.research.google.com/drive/1GGFtfbIZ2igx6gKjMGZXHP8VJ-__kYIm?usp=sharing

Run Locally

Clone the project

  git clone https://github.com/ty-dotcom/Grid-Monitoring-System.git

Download and setup PostgreSQL and TimeScaleDB

https://www.postgresql.org https://timescaledb.org

1. Copy backend/.env.example to backend/.env:

  cp backend/.env.example backend/.env

2. Edit your backend/.env and add your PostgreSQL credentials.

3. Ensure grid_db database exists and TimeScaleDB is installed, afterwhich run

  SELECT create_hypertable('grid_metrics', 'time', chunk_time_interval => INTERVAL '5 mins');

Using separate terminals/IDE

Go to the frontend directory

  cd Grid-Monitoring-System/frontend

Install dependencies

  npm install

Go to the backend directory

  cd Grid-Monitoring-System/backend

Install python requirements

  pip install -r requirements-dev.txt

Otherwise run python code using an IDE. Start the Backend server (Using Command Line).

  python main.py

Start the Client server

  npm run start