SHAP-based interpretable sparse autoencoder-neural network framework for electric vehicle energy consumption forecasting

Accurate energy consumption (EC) prediction for different road conditions is critical for battery management, regenerative braking control, and range estimation in battery electric vehicles (BEVs). However, conventional models trained on raw telemetry data often fail to capture complex driving dynamics due to noise and redundancy in high-dimensional inputs. To address these issues, this study proposes a novel hybrid interpretable framework that combines the sparse autoencoder, neural networks, and SHAP (Shapley additive explanations) (SAE-NN-SHAP), which effectively enables denoising, features extraction, reduces dimensionality, captures complex nonlinear characteristics and interpretable EC estimation, utilizing a real-world BEV dataset from a downloadable dynamometer database. Firstly, the enhanced SAE compresses the input space into 25 compact yet essential and meaningful information-dense latent features. Secondly, this representation is then fed into a Bayesian-optimized NN regressor. Both instantaneous (kW) and trip-level (Wh/km) EV-EC were estimated using the meaningful latent features. Thirdly, the proposed SAE-NN achieves the highest R ² (0.991) and lowest root mean squared error (RMSE) (1.270) and mean absolute error (MAE) (0.930) among all evaluated models, reducing error by 96.67% relative to a conventional NN trained on raw features and by 14.19% relative to the strongest baseline; at the trip level, predicted and actual EC differ by less than 1% across unseen trips’ driving conditions. Finally, the innovative end-to-end SHAP-based interpretability method, combined with SAE-NN encoder weights and latent feature mapping, identifies the specific driving-dynamics factors influencing EC, including speed, regeneration, braking, torque, battery dynamics, etc. The findings provide physically meaningful explanations and actionable insights for an accurate EC and range estimation, proactive and intelligent energy management approach.