| 摘要: |
| 水路货运量需求受诸多因素影响，长江干线中游“645”工程实施后，航道通航条件得到了明显改善，为了更好分析工程实施后货运量变化趋势，提出一种新的水路货运量预测模型。首先，采用二次插值法和KNN反距离权重插值法解决高维面板数据中时间粒度不统一与缺失问题，利用层次聚类和SHAP值的可解释性综合筛选关键影响因素特征序列，降低预测模型输入数据的维度和规模，引入Halton低差异序列和准反射学习策略(QRBL)大幅提升金枪鱼群优化算法(TSO)的寻优效能，增强TSO算法对极限梯度提升(XGBoost)模型中决策树数量、决策树的深度、学习速率等决定模型拟合能力的超参组合寻优效果。计算结果表明，新模型预测精度显著优于对比模型，其平均绝对百分比误差、平均绝对值误差、均方根误差分别：2.05%、102.2335、145.0957，可更好的适用于多特征影响因素下的水路货运量预测研究。 |
| 关键词: SHAP Halton低差异序列 QRBL TSO XGBoost |
| DOI： |
| 分类号: |
| 基金项目:国家重点研发计划项目(2018YFB1600405)；国家自然科学基金(52271367) |
|
| Waterway freight volume prediction based on SHAP and multi-strategy optimization TSO-XGBoost model |
|
Wen Quan1, YU Yuhuan2, Zhuang Shangde2, Mou Junmin3
|
|
1.Wuhan Vocational College of Software and Engineering(Wuhan Open University);2.Changjiang Waterway Bureau;3.Wuhan University of Technology
|
| Abstract: |
| Waterway freight volume demand is affected by many factors. After the implementation of the "645" project in the middle reaches of the Changjiang River main line, waterway navigation conditions have been significantly improved. In order to better analyze the trend of freight volume after the implementation of the project, a new waterway freight volume forecast model is proposed. Firstly, quadratic interpolation method and KNN inverse distance weight interpolation method were used to solve the problem of time granularity disunity and missing in high-dimensional panel data. Hierarchical clustering and interpretability of SHAP values were used to comprehensively screen feature sequences of key influencing factors to reduce the dimension and scale of input data in prediction model. Halton low difference sequence and quasi-reflection-based learning (QRBL) are introduced to greatly improve the optimization efficiency of tuna swarm optimization (TSO), and enhance the optimization effect of TSO algorithm on the number of decision trees, the depth of decision trees and the learning rate, which determine the fitting ability of the model in XGBoost model. The results show that the prediction accuracy of the new model is significantly better than that of the comparison models, and mean absolute percentage error, mean absolute error and mean absolute error are 2.05%, 102.2335 and 145.0957, respectively, which can be better applied to the research of waterway freight volume prediction under multi-feature factors. |
| Key words: SHAP Halton low difference sequence QRBL TSO XGBoost |
