Time Series Forecasting of Electricity Load Using Hybrid CNN-BiLSTM with an Attention Approach: A Case Study in Bali, Indonesia

The use of electrical energy continues to increase from time to time, in line with the promotion of green energy. Prediction of the use of electricity load is crucial for increasing electricity production efficiency, especially for electricity companies. One way of approaching electrical load prediction is with a machine learning model. This study modeled short-term time series forecasting of electricity load using the CNN-BiLSTM hybrid model with an attention mechanism to calculate short-term predictions, namely 1, 3, and 7 days ahead. The study aims to see the effects of additional attention mechanisms in the hybrid model CNN-BiLTSM with the resulting accuracy that the model gets. Furthermore, we also compare the CNN-BiLSTM hybrid model with attention to the BiLSTM and CNN-BiLSTM models. As a case study, this study used data on the electricity load in Bali, Indonesia. We also compare the experimental results to show that the CNN-BiLSTM with Attention model produces the best accuracy, with a root mean square error of 13.676 and a correlation coefficient 0.993 compared to the other two. These experiments have shown that improving predictions of electricity usage can be achieved by combining the original BiLSTM model with both a CNN layer and an attention mechanism. While the tests conducted primarily focused on short-term forecasting, the CNN-BiLSTM model with an attention mechanism shows potential for further exploration in the realm of intermediate and long-term forecasting.