Analysis of power system load forecasting based on neural networks

Abstract

In this research, our main goal was to improve power load forecasting accuracy by considering the impact of meteorological factors on the total power of the electrical system, examining existing load data, local weather, wind direction, and other parameters affecting total power load. We divided the data from the past three years into a training dataset, comprising 75% of the data, and a testing dataset with the remaining 25%. We employed a basic machine learning technique (Support Vector Machine) and three distinct neural network approaches (Artificial Neural Network, Convolutional Neural Network, and Long-Short Term Memory Network) to develop analytical models. Through experimentation, the LSTM model achieved a loss value of 0.0034 and required 1426.78 seconds of training time across 100 epochs. Considering the time expense and model complexity, we chose the LSTM model to forecast power load at 15-minute intervals for the subsequent ten days, achieving a satisfactory prediction and fitting outcome. Our results suggest that the LSTM model is a promising method for optimizing performance and reliability in electrical power systems.