ELECTRICITY DEMAND FORECASTING USING HYBRID MACHINE LEARNING MODEL

Abstract

Accurately predicting electricity consumption is essential for maintaining the stability and effectiveness of power networks in an era of swift technical breakthroughs and rising global energy demand. In order to estimate power demand, this thesis suggests a hybrid machine learning framework that combines Random Forest (RF) and Convolutional Neural Networks (CNNs). The CNN-RF model aims to improve prediction accuracy and robustness by utilising the feature extraction skills of CNNs and the decision-making characteristics of RF. CNN applications are made easier by the study's use of the Gramian Angular Field (GAF) approach, which converts time-series data into image representations. Based on Pennsylvania-New Jersey-Maryland East (PJME) Interconnection dataset experimental assessments, the CNN-RF model outperforms conventional approaches such as XG-Boost and standalone CNN models in terms of forecast accuracy Metrics like Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and Mean Squared Error (MSE) are used to validate the model's performance. Our results demonstrate that the hybrid CNN-RF approach effectively captures complex spatiotemporal patterns in electricity demand data, offering a reliable tool for energy resource management. The thesis also explores the implications of input image size on model performance and suggests future enhancements through the integration of attention mechanisms to optimize data preprocessing.