基于MI特征选择的车辆能耗高精度预测方法

High-Precision Vehicle Energy Consumption Prediction using Mutual Information Feature Selection

摘要:

近年来，机器学习方法在车辆实时能耗预测方面得到了广泛应用，但实车采集数据中存在的精度不足、字段缺失以及多重共线性等问题，尤其是同款车型中驾驶工况和驾驶者行为存在显著差异，限制了能耗预测准确性和泛化能力的进一步提升。为此，本文系统考虑特征冗余度、数据平衡性、货运趟次、运输能力、路段拥挤程度和司机驾驶时长等因素，使用交互信息（MI）方法选择关键特征，并构建司机特征画像作为独立特征，进而结合XGBoost、RF和MLP等机器学习方法提出一种基于MI特征选择的能耗高精度预测方法，然后基于120辆轻型卡车的T-BOX采集数据进行实例验证。结果表明，本文提出的预测方法能够显著提高不同驾驶行为和驾驶工况下的能耗预测精度，研究成果可为开发预测轻卡能耗的通用模型提供参考。

Abstract:

In recent years， machine learning methods have been widely adopted for real-time vehicle energy consumption predictions. However， the accuracy and generalizability of these predictions are often hindered by challenges such as data imprecision， missing fields， multicollinearity， and substantial difference in driving conditions and driver behaviors among identical vehicle models. To address these issues， this study systematically considers factors such as feature redundancy， data balance， freight trip frequency， transport capacity， traffic congestion and driving duration. Subsequently， an energy consumption prediction model with high precision is developed using a combination of machine learning methods such as XGBoost， Random Forest （RF）， and Multilayer Perceptron （MLP）. The model utilizes key features selected through the Mutual Information （MI） method， along with a constructed driver profile that captures characteristic behaviors as an independent feature. The proposed method is validated using T-BOX data collected from 120 light trucks. Experimental results indicate that the prediction method significantly enhances the prediction accuracy of energy consumption under various driving behaviors and conditions. This research contributes to the development of models with high precision in estimating the fuel consumption of light trucks.