Short-term Power Load Forecasting for a 33/11 KV Sub-Station by Utilizing Attention-Based Hybrid Deep Learning Architectures

Estimating electric power load at substations is a fundamental task for system operators, as it is essential for the reliable and optimal operation of the power system. Effective load forecasting is critical for optimal power generation, as precise predictions facilitate the economical use of electrical infrastructure. The primary objective of this study is to develop advanced deep learning (DL) attention-based models aimed at improving the accuracy of short-term electric power load forecasting at substations. This enhancement is essential for ensuring the reliable and efficient operation of power systems. To accomplish this objective, a comprehensive evaluation of various machine learning (ML) and deep learning (DL) architectures was conducted. This evaluation included the following models: Autoregressive Integrated Moving Average (ARIMA), Multi-Layer Perceptron (MLP), Ran-dom Forest (RF), Gradient Boosting (GB), Long Short-Term Memory (LSTM) networks with Atten-tion mechanisms, LSTM-Convolutional Neural Network (CNN) Attention, etc. These models applied to hourly estimated energy consumption data (in kilowatts) sourced from the 33/11 kV sub-station in Telangana, India. The performance of these models measured using several key metrics, including Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R-squared (R²). The most important result is that the CNN-BiLSTM attention model significantly outperforms the other models, achieving an MSE of 0.0079, an RMSE of 0.0889, and an R² value of 0.8547. that under-scores that the CNN-BiLSTM attention model represents an effective and practical tool for accurate power load forecasting. This capability not only enables the economical utilization of electrical infrastructure but also supports reliable, data-driven decision-making processes within power system operations.