基于NeuralProphet-LSTM组合模型的港口货物吞吐量预测

Prediction of port cargo throughput using NeuralProphet-LSTM combination model

为进一步提高货物吞吐量预测准确性,提出基于NeuralProphet时间序列模型与长短期记忆(LSTM)神经网络的组合预测模型。首先利用NeuralProphet模型对港口货物吞吐量数据进行训练得到预测值并计算残差序列,然后对残差数据建立LSTM神经网络模型进行预报修正,重构得到最终的预测值。以上海港、厦门港的月度货物吞吐量数据为样本展开试验,结果表明,该模型能够有效地解决数据异常波动造成的预测结果误差大、预测效果不稳定等问题;相比于传统单一模型与LSTM-支持向量机(SVM)、Bi-LSTM等组合模型,NeuralProphet-LSTM模型预测精度更高,可帮助港航企业及时调整规划决策与经营策略。

Conventional method has not been successful in predicting cargo throughput of ports due to the great fluctuation caused by complex external factors.In order to improve the accuracy of cargo throughput prediction for ports,the combination of NeuralProphet time series model and LSMT neural network is introduced.NeuralProphet model is trained to produce,according to input cargo throughput data,a preliminary prediction with a residual sequence.The LSMT model is used to correct the preliminary prediction and output final prediction.The combination model is verified through process the monthly throughput data from Shanghai Port and Xiamen Port.The advantage of the combination model in handling abnormal fluctuation over other configurations is demonstrated.