Abdul-Pandev/DumsorData

Predicting Ghana's power outages with real data, machine learning, and solutions that actually matter. # DumsorData 🔦 > *The data already knows where the lights will go out next.* Ghana has lived with dumsor for decades. DumsorData turns that frustration into insight. Using real sensor data from nLine GridWatch spanning 2018 to 2024, this project applies machine learning to predict power outages across Accra, identify the chronic problem sites, and recommend targeted solutions — because the data already knows where the lights will go out next. --- ## Why This Project Exists If you have ever lived in Ghana, you know dumsor. You know the sound of the neighbours generator starting at 11pm. You know the feeling of losing hours of work because the power cut mid-session. You know the trader who lost a week of stock because the freezer went off. You know the student who could not study. The hospital that scrambled. The family that just sat in the dark and waited. What most people do not know is that the data for this problem already exists. Real sensors. Real readings. Real patterns. Sitting in a public database, waiting for someone to do something with it. This project is that something. --- ## What DumsorData Does DumsorData is a complete end-to-end machine learning project that: - Downloads and merges real Ghanaian electricity outage data from the nLine GridWatch public dataset - Cleans, engineers features, and handles class imbalance using SMOTE - Trains and tunes three classification models to predict weekly power outages at sensor sites - Produces a full technical notebook for the data science community - Produces a plain language report with six concrete recommendations for ECG, government, and communities --- ## The Data The dataset comes from **nLine GridWatch**, a project that deployed over 1,200 physical sensors across 14 districts in Accra's ECG (Electricity Company of Ghana) service area. | Property | Detail | |---|---| | Source | nLine GridWatch — nline.io | | Coverage | June 2018 to August 2024 | | Sensor Sites | 445 unique locatio …