Cross-Validated Neural Network Optimization for Explainable Energy Prediction in Industrial Mobile Robots

Accurate energy prediction is essential for energy-aware planning and navigation of autonomous mobile robots (AMRs). This study investigates whether a compact feed-forward neural network (FFNN) can predict relative energy consumption from operational variables with high accuracy and interpretability. Using a curated dataset of controlled translation and rotation trials on a KUKA KMP 1500P, energy demand is expressed as the per-trial reduction in battery state of charge (SoC), defined as N1%. For unit-free reporting, it is also considered the normalized SoC consumed per trial, defined as E=1/N1%. Model development followed a two-stage optimization pipeline, (i) systematic feature-subset screening and (ii) cross-validated architecture and regularization search with early stopping, assessed by a composite of MSE, MAE, R2, and the 68th-percentile absolute error (∆X68) as the prediction precision index. The selected FFNN (ReLU multilayer perceptron with L2 weight decay) achieved strong generalization on the independent test set (MAE = 0.9954, MSE = 4.5512, R2 = 0.9795, ∆X68 = 0.0193). Post hoc explainability methods (SHAP and input perturbation) identified angular velocity and linear acceleration as the dominant predictors, with payload mass exerting secondary effects. These results demonstrate that a compact, regularized FFNN provides accurate, repeatable, and interpretable energy predictions suitable for integration into digital twin platforms and downstream industrial scheduling.