An Optimized Port Operation Efficiency Prediction Model Based on ESN and LSTM

Abstract

With the in-depth digital transformation of the global shipping industry, the accurate prediction of smart port operation efficiency has become a key factor in enhancing the competitiveness of international supply chains. Aiming at the limitations of traditional models in handling nonlinear dynamics and noisy multi-scale data in port operations, this paper proposes a dual-model optimization framework based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Echo State Network (ESN), and Bidirectional Attention-Enhanced Long Short-Term Memory Network (Bi-ALSTM). At the macro level, CEEMDAN is used to decompose the multi-scale features of port efficiency signals, and a lightweight prediction model is constructed in combination with ESN. At the micro level, a bidirectional attention mechanism is introduced to improve LSTM, enhancing the ability to model temporal dependencies. Experimental results show that the training time of CEEMDAN-ESN is only 0.35 seconds, demonstrating a significant real-time advantage; the Mean Absolute Error (MAE) and Mean Squared Error (MSE) of Bi-ALSTM on the test set are 123.47 and 25475.36, respectively, which are 18.10% and 30.50% higher than those of the traditional Recurrent Neural Network (RNN) model. These results verify the comprehensive advantages of the proposed models in terms of accuracy and efficiency. This study provides an interpretable quantitative tool for dynamic scheduling and risk early warning of smart ports, and helps transform port management towards an automated and predictive model.