~10M
Raw meter-reading records consolidated
Back to projects
01 02 03 04 05 06 01 02 03 04 05
Government Internship Project
Electricity Consumption Modeling
Large-Scale Energy Usage Analysis in R
A large-scale data-analysis and baseline-modeling project completed during a summer internship at the Digital Ministry of the Republic of Uzbekistan. The project examined electricity usage patterns and prepared a reproducible foundation for future energy-demand forecasting.
Electricity Consumption ModelingGovernment Internship Project
427,900
Monthly consumer observations prepared for modeling
3 years
Historical electricity-consumption data analyzed
Objective
Building a foundation for future energy forecasting
The project began with electricity meter readings, confirmed payment records, and monthly account-balance files. The primary analytical target was monthly electricity consumption.
The immediate objective was to organize the historical data, understand usage patterns, create a modeling-ready dataset, and evaluate an interpretable regression baseline before developing more advanced forecasting methods.
Analytical workflow
From raw meter readings to temporal model evaluation
Data consolidation
Combined ten reading files into one time-ordered dataset and standardized consumer identifiers, dates, and meter values.
Consumption calculation
Converted cumulative meter readings into actual consumption by calculating the difference between consecutive readings for each consumer.
Monthly aggregation
Aggregated daily readings into consumer-level monthly consumption records to support trend analysis and forecasting.
Exploratory analysis
Examined seasonality, yearly distributions, individual consumer patterns, concentration of high-use consumers, and possible outliers.
Feature engineering
Created seasonal flags, prior-month consumption, consumption change, log-transformed usage, and high-usage indicators.
Temporal modeling
Trained a linear-regression baseline using 2022–2023 data and evaluated it on the 2024 observation period.
Temporal validation
Testing the model on a later time period
Training period
2022–2023
279,278 monthly observations
Testing period
2024
148,622 monthly observations
Key findings
Seasonality, skewed usage, and consumer heterogeneity
Overall monthly electricity consumption showed recurring seasonal variation and a modest upward trend.
Consumption tended to increase during colder months, supporting the inclusion of seasonal features.
Daily-consumption values were strongly right-skewed, with a relatively small number of very high-use consumers.
Consumer behavior varied substantially, and several extreme values required additional investigation.
The baseline model captured broad patterns but produced substantial relative error for low-use and irregular consumers.
Baseline evaluation
Measuring prediction error on 2024 data
MAE
6,590.74 kWh
RMSE
10,020.15 kWh
MAPE
63.62%
The MAPE remained high, indicating that predictions were less reliable for consumers with very low usage, irregular behavior, or extreme observations.
Modeling lesson
Evaluating features as carefully as model performance
The baseline model helped expose several issues that would need to be addressed before production forecasting, including outliers, zero-consumption records, missing lag values, and differences between household and high-use consumers.
One engineered variable, average daily consumption, was derived directly from monthly total consumption. Because the monthly total was also the prediction target, this created target leakage and made the apparent model fit overly optimistic.
A stronger next iteration would remove target-derived predictors, use only information available before the forecast period, segment different consumer types, and compare time-series or tree-based models using rolling temporal validation.
Technology
Tools used for data preparation, analysis, and modeling
Rtidyversedplyrlubridateggplot2Linear RegressionFeature EngineeringTemporal Validation