Comparing MLP and 1D-CNN Architectures for Accurate RUL Forecasting in Lithium Batteries

Abstract

Accurately forecasting the Remaining Useful Life (RUL) of lithium-ion batteries is critical for optimizing battery management and ensuring operational reliability. This study compares the performance of two deep learning architectures—a Multilayer Perceptron (MLP) and a one-dimensional Convolutional Neural Network (1D-CNN)—in predicting RUL using datasets from CALCE batteries B35, B36, and B37. Data preprocessing involved outlier removal, missing value handling, and feature normalization, with key features extracted including Resistance, Constant Voltage Charging Time (CVCT), and Constant Current Charging Time (CCCT). Correlation analyses confirmed strong relationships between these features and RUL. Both models were trained and validated on preprocessed data, and their predictive accuracies were assessed using Root Mean Square Error (RMSE) and coefficient of determination (R2). Results indicated that while both architectures effectively captured battery degradation patterns, the MLP consistently outperformed the 1D-CNN, achieving on average 5% lower RMSE and 1.5% higher R2 across all tested batteries. These findings suggest that simpler fully connected networks may suffice for this forecasting task under the given feature set and preprocessing conditions. This work provides valuable insights into neural network model selection for battery health prognostics, guiding the development of efficient and accurate predictive maintenance strategies.