Electricity Bill Prediction Based on a Particle Swarm Optimized Multilayer Perceptron Model

Accurate household electricity bill prediction enables better budgeting for consumers and data-driven planning for utilities. This study develops and benchmarks five deep learning models on a publicly available Indian household electricity bill dataset that combines appliance usage and socio-demographic attributes. We propose a Particle Swarm Optimized Multilayer Perceptron (PSO-MLP) model that tunes network depth, width, learning rate, and regularization via Particle Swarm Optimization (PSO), and compare it against plain Multilayer Perceptron (MLP), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Recurrent Neural Network (RNN) architectures. The pipeline includes robust preprocessing (median imputation, scaling, and one-hot encoding), leakage-safe training/testing, and a comprehensive evaluation suite comprising Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Squared Error (MSE), coefficient of determination (R²), and Median Absolute Error (MedAE). Results show a near-deterministic fit: PSO-MLP achieves MAE = 10.22, RMSE = 12.99, MSE = 168.92, R² = 0.9998, and MedAE = 8.43; a plain MLP attains MAE = 10.22 with a similar R², whereas recurrent models provide no advantage on this non-sequential, tabular task (RNN MAE = 23.03). Error distributions confirm stable performance across the bill range with minimal bias. These findings indicate that carefully regularized feed-forward models—augmented with principled hyperparameter optimization—suffice to model household bills with very high fidelity, whereas more complex sequence models are unnecessary. The proposed framework offers a strong baseline for tariffaware extensions and deployment-grade forecasting in Indian residential settings.