Electrical Load Forecasting using Long Short-Term Memory (LSTM)

Abstract

Accurate electric load forecasting is essential for effective energy management and optimization. This project uses Long Short-Term Memory networks, a type of Recurrent Neural Network, to predict electric load due to their ability to capture long-term dependencies in time-series data. Features such as weather and holidays affect load consumption. This project aims to determine whether the number of features and the number of Long Short-Term Memory layers impact prediction accuracy. The project develops and evaluates 1-layer, 2-layer, and 3-layer models for forecasting. The two years of data used for this study is sourced from Kaggle, provided by a power supply company in Johor, Malaysia. Additional weather data is collected from the Weather Underground website, and holiday information includes public holidays and weekends. The Long Short-Term Memory models are trained using electric load data combined with features such as time, holiday, temperature, and humidity. Model performance is evaluated using metrics like Mean Absolute Error, Root Mean Squared Error, Mean Absolute Percentage Error, and Mean Squared Error. This project compares results across the three LSTM architectures and discusses the effects of data modifications, model architecture, and the training/validation process on forecasting outcomes. The findings show that the 1-layer Long Short-Term Memory model achieves the best accuracy, with a Mean Absolute Percentage Error of 15%, outperforming the 2-layer and 3-layer models, which achieve 16% and 18% Mean Absolute Percentage Error, respectively. These results can help utility companies optimize electricity generation, plan equipment maintenance, and improve resource procurement strategies.