Predictive Modeling of Electricity Load Demand Forecasting Using the CNN-BiLSTM method based on Peak Load in Household Sector Consumers

Abstract

Accurate short-term electricity load forecasting is essential for ensuring reliable energy
management and maintaining power system stability, particularly in the household sector where
electricity consumption exhibits highly dynamic and nonlinear patterns. Conventional forecasting
methods often have limited capability in capturing these complex temporal characteristics.
Therefore, this study proposes a hybrid Convolutional Neural Network–Bidirectional Long ShortTerm Memory (CNN-BiLSTM) model to forecast 24-hour ahead household electricity demand based
on peak load data collected from Mojowarno District, Jombang Regency, Indonesia. The dataset
consists of hourly electricity consumption records from January 2024 to January 2025 and was
preprocessed through smoothing, outlier handling, and normalization before model training. The
proposed model combines CNN for automatic spatial feature extraction and BiLSTM for learning
bidirectional temporal dependencies. Experimental results demonstrate excellent forecasting
performance with a Test Loss of 0.0024, Test MAE of 0.0562, Test RMSE of 0.0699, MAE of 3.4146
kW, RMSE of 4.2470 kW, and an R² value of 0.9889. These findings indicate that the proposed CNNBiLSTM model effectively captures household electricity consumption patterns and provides
accurate short-term peak load forecasting, making it a promising approach for supporting energy
management and electricity distribution planning