File size: 3,287 Bytes
58216f5 78a2cc6 58216f5 78a2cc6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 | ---
library_name: scikit-learn
tags:
- scikit-learn
- regression
- time-series-forecasting
- electricity-consumption
- pakistan
---
# ⚡ BijliCast
## Overview
BijliCast is a weather-aware household electricity
consumption forecasting system developed for Pakistan.
It uses historical household electricity consumption,
weather conditions, calendar information and household
characteristics to predict next-hour electricity usage.
The prediction target is `next_hour_energy_kwh`.
## Model Information
| Property | Value |
|---|---|
| Model | HistGradientBoostingRegressor |
| Framework | Scikit-learn |
| Task | Time-series regression |
| Artifact | deployment_model.joblib |
| Model size | Approximately 1.4 MB |
| Input features | 42 raw features |
The saved artifact includes the fitted preprocessing
pipeline and the trained estimator.
## Dataset
BijliCast combines two independent public data sources:
1. LUMS REWD-P household electricity dataset.
2. Open-Meteo ERA5 historical weather data.
Dataset characteristics:
- 59 households
- Six Pakistani cities
- 515,786 hourly electricity records
The weather information is joined with household
electricity measurements using temporal and geographic
alignment.
## Methodology
The ML lifecycle includes:
- Exploratory Data Analysis
- Missing-value handling
- Feature engineering
- Lag and rolling statistics
- Chronological train/validation/test splitting
- Scikit-learn preprocessing Pipeline
- Baseline comparisons
- Multiple regression algorithms
- Five-fold TimeSeriesSplit cross-validation
- Randomized hyperparameter optimization
The final deployment pipeline was fitted using
training and validation data. The held-out test set
was excluded from model fitting and selection.
## Model Selection
Random Forest achieved a validation MAE of 0.3216 kWh,
while HistGradientBoosting achieved 0.3273 kWh.
Their cross-validation variation indicated practically
comparable performance.
HistGradientBoosting was selected for deployment
because it offered comparable predictive quality
with a substantially smaller serialized artifact.
Random Forest: approximately 270 MB
HistGradientBoosting: approximately 1.4 MB
This made HGB considerably more practical for
GitHub-based and lightweight deployment.
## Evaluation Results
Held-out test results:
| Metric | Result |
|---|---:|
| MAE | 0.2499 kWh |
| RMSE | 0.4273 kWh |
| R² | 0.7291 |
These are final evaluation results, not the
criteria used to select the model.
## Prediction Interval
The application provides a 90% prediction interval
using residual-based offsets estimated from
out-of-sample validation predictions.
## Intended Use
- Academic research
- Household electricity consumption analysis
- Short-term residential load forecasting
- Educational machine learning demonstrations
## Limitations
- The dataset covers 59 households.
- Coverage is limited to six Pakistani cities.
- The demo uses historical replay data rather
than live smart-meter readings.
- The model is not intended for electricity billing
or grid operational decisions.
## Repository
Full source code, notebooks, reports and deployment:
https://github.com/muhammad-umar-218980/BijliCast
## Data Sources
- https://rewdp.lums.edu.pk/
- https://open-meteo.com/ |