Performance Comparison of TPE, ASHA, and Genetic Algorithm for Hyperparameter Optimization in Transformer-Based Water Consumption Forecasting

Abstract

This study presents a comparative analysis of hyperparameter optimization (HPO) techniques to enhance the performance of Transformer-based models in water consumption time series forecasting. Accurate demand prediction is essential for efficient water resource management, yet Transformer models are highly sensitive to hyperparameter configurations, making optimization a critical challenge. In this study, three widely used HPO methods—Tree-structured Parzen Estimator (TPE), Asynchronous Successive Halving Algorithm (ASHA), and Genetic Algorithm (GA)—are evaluated within a unified experimental framework. The experiments are conducted on a real-world multivariate water consumption dataset with 46,032 observations, incorporating temporal feature engineering and feature scaling to improve model stability. Performance is evaluated using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) across multiple forecasting horizons (H1, H6, and H24). The results show that ASHA consistently achieves the best performance, reducing MAE by up to 13.4% compared to GA and 7.8% compared to TPE at the long-term horizon. Additionally, ASHA demonstrates more consistent performance across forecasting horizons, indicating improved generalization capability in multi-step prediction. These findings highlight the effectiveness of ASHA as a computationally efficient approach for optimizing Transformer-based models in time series forecasting applications.