Short-Term Load Forecasting In Thimphu and Phuentsholing Regions Using Machine Learning and Deep Learning Techniques

Abstract

Accurate short-term load forecasting is essential for efficient power system operation, energy
management, and electricity pricing. Traditional statistical methods, such as seasonal autoregressive
integrated moving averages with exogenous variables (SARIMAX), often fail to capture the
intricate and dynamic patterns of electricity demand. Addressing the knowledge gap in developing
countries, particularly Bhutan, this research explores advanced machine learning and deep learning
techniques to enhance short-term load forecasting (STLF) accuracy in the Thimphu and
Phuentsholing regions of Bhutan, characterized by unique electricity demand patterns due to
population growth, industrial and commercial activities, and supply constraints. We evaluated
SARIMAX, Support Vector Regression (SVR), Long Short-Term Memory Networks (LSTM),
Convolutional Neural Networks (CNN), and hybrid CNN-LSTM architectures. On single-step
STLF, we analyzed day-ahead aggregated load forecasts and 1-hour-ahead load forecasts based on
historical load data over five years (2018-2022) for both Thimphu and Phuentsholing regions. In
day-ahead aggregated load forecasting, the hybrid CNN-LSTM outperformed all other models with
Mean Absolute Percentage Error (MAPE) values 2.332 ± 0.075% for Thimphu and 3.216 ±

0.036% for Phuentsholing, while also achieving the best MSE, RMSE, and R2 metrics. For 1-hour-
ahead forecasting, the CNN model achieved the lowest MAPE of 3.224 ± 0.06% in Thimphu and

the hybrid CNN-LSTM model achieved a best MAPE of 3.687 ± 0.027% for Phuentsholing.
Careful preprocessing, optimal feature engineering, and hyperparameter tuning were performed for
all forecasting types. The findings demonstrate that data-driven approaches significantly enhance
forecasting accuracy, providing valuable insights for energy planners to manage resources and
maintain the power grid, preventing blackouts and other disruptions.