A Data-Driven Approach to Short-Term Load Forecasting using LSTM Neural Networks for Grid Operations

Electricity is delivered to consumers through a complex, interconnected power system involving generation, transmission, and distribution. Demand for electricity is highly dynamic, influenced by temporal, seasonal, demographic, economic, and environmental factors. Accurate load forecasting is essential for grid reliability, operational efficiency, infrastructure planning, and cost optimization. Forecasting horizons are typically classified into short-term, medium-term, and longterm, each serving different strategic and operational objectives. Traditional statistical methods, while effective in stable conditions, often fall short in capturing non-linear and stochastic demand patterns. As a result, machine learning and deep learning models have gained prominence in recent years. This study focuses on deep learning approaches, specifically comparing Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNN) for short-term load forecasting. LSTM, a type of recurrent neural network, is adept at learning long-term temporal dependencies in sequential data, making it suitable for time series forecasting. In contrast, CNNs, though originally designed for image processing, are effective in capturing local patterns in time series data through 1D convolutions. This paper evaluates the forecasting performance, pattern recognition capabilities, and contextual suitability of both models for power system operations. The findings aim to support utilities in selecting the most appropriate forecasting approach under varying system conditions and requirements.