Short-Term Electricity Load Forecasting using Deep Learning Models for the Renewable-Integrated Power System

Abstract

Short-term electricity load forecasting is critical for reliable power system operations and economic planning. This study evaluates deep-learning models based on gated recurrent units (GRUs) to predict day-ahead electricity demand from hourly system load data obtained from the California Independent System Operator (CAISO). The models used the hour of the day, day of the week, and lagged loads at intervals of 24, 48, 72, 168, and 336 hours as input features. Here, two GRU architectures have been used for the analysis. Firstly, a base GRU model with a single GRU layer and secondly, a stacked GRU model with two GRU layers. Both GRU architectures were trained with the same hyperparameters in MATLAB R2025b, and their performance was assessed using the mean absolute percentage error (MAPE), mean absolute error (MAE), and coefficient of determination (R2). On the global test set, the stacked GRU achieved a MAPE of 3.6623%, an MAE of 982.8549 MW, and an R2 of 0.91427, outperforming the base GRU, which yielded a MAPE of 3.9325%, an MAE of 1049.4558 MW, and an R2 of 0.90669. A 14-day span analysis further showed consistent performance gains in most 14-day intervals. The results indicate that the stacked GRU model outperforms the base GRU model on the load dataset, leading to improved day-ahead forecasts and providing a useful resource for the system operators.