Optimization of Variable Combinations for Household Electricity Consumption Prediction Using a Multivariate Time Series Machine Learning Approach

Abstract

Accurate household electricity consumption prediction is vital for effective energy planning in Indonesia, a nation facing rapid economic growth and technological advancements. Inaccurate predictions can lead to inefficiencies in resource allocation and energy shortages. Traditional methods like ARIMA struggle with non-linear patterns, long-term dependencies, and multivariate relationships critical in understanding electricity consumption dynamics. To address these challenges, this study employs the Long Short-Term Memory (LSTM) algorithm with a multivariate time series approach, chosen for its ability to capture complex patterns and long-term trends. The dataset comprises monthly electricity consumption data (2004–2023) from PT PLN, enriched with macroeconomic and environmental variables like Household Consumption GDP, inflation, and average temperature. The Denton-Chollete method was used to transform quarterly GDP data into monthly intervals, and correlation analysis identified Household Consumption GDP (r=0.98) and Power Contract Additions (r=0.64) as significant predictors. Testing 63 feature combinations, the best (Power Contract Additions, Household Consumption GDP, and Household Electricity Consumption) achieved a Mean Absolute Percentage Error (MAPE) of 3.54%. These results highlight LSTM's superiority in handling dynamic and complex electricity consumption patterns and provide a robust predictive tool for PT PLN. This study underscores the importance of exploring additional variables and advanced optimisation techniques to enhance predictive accuracy further.