Enhanced Load Forecasting Using CNN-BiLSTM Models in University Buildings with Solar PV

• Published in: SSRG International Journal of Electrical and Electronics Engineering
• Main Author: Ahmad M.Z.; Dahlan N.Y.; Yasin Z.M.
• Format: Article
• Language: English
• Published: Seventh Sense Research Group 2024
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85209214424&doi=10.14445%2f23488379%2fIJEEE-V11I10P107&partnerID=40&md5=8d59ca0913fee5e3c97266115c2bfbef

Energy management is an imperative practice involving the detailed monitoring, regulation, and optimization of energy consumption within various domains aimed at conserving resources and controlling energy costs. The escalating demand for electricity and integrating renewable energy sources has brought forth an array of complexities that challenge energy management efforts. As a result, the need to enhance the precision of load forecasting has surged in importance, attracting significant attention from researchers and organizations alike. Traditional time series models have their own limitations. These models rely on the assumption of linear relationships and stationary time series data, thereby potentially falling short of capturing the intricate, non-linear variations often present in energy consumption patterns. This limitation necessitates the exploration of more advanced and adaptable forecasting techniques. Campus buildings present some challenges for load forecasting. These challenges arise from the dynamic and ever-changing load patterns within educational institutions, which can fluctuate significantly based on various factors such as lecture schedules, semester breaks, and special occasions. This study introduces an innovative hybrid model called CNN-BiLSTM, which integrates Convolutional Neural Networks (CNN) with Bidirectional Long Short-Term Memory (BiLSTM) models to address the complexities of load patterns to produce accurate forecasts. The proposed model is thoroughly benchmarked against traditional Artificial Neural Networks (ANN) and BiLSTM models. Load data from the UiTM Permatang Pauh campus building, which encompasses 343 days of data collected at 30-minute intervals, a total of 16,464 data points for analysis. Leveraging this load data, comprehensive feature engineering was conducted, leading to the generation of categorical data such as hour, calendar attributes, and semester status. The CNN-BiLSTM model outperforms its counterparts, achieving a remarkable Mean Absolute Percentage Error (MAPE) of 6.9%. Therefore, as demonstrated through rigorous benchmarking, the model's superior performance highlights its potential significance for improving energy management in educational institutions and other domains with similar load complexity. © 2024 Seventh Sense Research Group®.