Bidirectional Long Short-Term Memory (Bi-LSTM) Hourly Energy Forecasting

Department of Electrical Engineering and Informatics, Faculty of Engineering, Universitas Negeri Malang, Malang 65145, Indonesia

^* Corresponding author: aji.prasetya.ft@um.ac.id

Abstract

The growing demand for energy, especially in urban and densely populated areas, has driven the need for smarter and more efficient approaches to energy resource management. One of the main challenges in energy management is fluctuations in energy demand and production. To overcome this challenge, accurate and careful forecasting of hourly energy fluctuations is required. One method that has proven effective in time series forecasting is using deep learning. The research phase uses the CRISP-DM data mining methodology as a common problem solver for business and research. The scenarios tested in the study used 5 attribute selection scenarios based on correlation values based on target attributes and 2 normalization scenarios. Then, the deep learning model used is Bi-LSTM with hyperparameter tuning grid search. Performance measurement evaluation is performed with MAPE, RMSE, and R2. Based on the tests conducted, it was found that the Bi-LSTM model produced the best MAPE of 7.7256%. RMSE of 0.1234. and R2 of 0.6151 at min-max normalization. In comparison, the results on the z-score normalization are lower with the best MAPE value produced at 10.5525%. RMSE of 0.7627. and R2 of 0.4186.