Research On Container Throughput Time Series Forecasting Methods

With the strategy of "The Belt and Road" attracts the intense attention in international society, the twenty-first Century Maritime Silk Road is becoming a crucial component of this strategy. Therefore, the international shipping industry has entered a new stage of development and containerization is gradually becoming the symbol of ocean transportation. Thus, an accurate forecast of the future port throughput is significant to decision maker in port for planning and managing their future development. This is because the operation of the port without an appropriate strategic planning, foresight for the future could lead to the dilemma of traffic jams or idle cost. In order to provide decision support for port appropriate operations and increase productivity and the efficiency of the port, a time series forecasting model of the future port throughput becomes a rather pioneering and indispensable solution.Firstly, we analyzed the characteristics of container time series and found it is a non-stationary series composed of linear and non-linear component. Also, the four features are founded: a long-term trend, seasonal trend, cycle fluctuation trend, irregular trend, which constitute the non-stationarity.Secondly, depending on the characteristics of time series and the modeling framework of linear and nonlinear forecasting model, two non-homologous forecasting models of seasonal autoregressive integrated moving averages(SARIMA), which is a high fitting degree on the linear component, and artificial neural networks(ANNs), which has a flexible fitting degree on the non-linear component, are integrated. However, conventional hybrids are proposed based on two types of assumption:(1)the final forecasting value is the simple average of forecasting values of individual linear and nonlinear forecasting models;(2)the relationship between linear component and nonlinear component in time series is additive, which are demonstrated not reasonable.To avoid the limitation of these two assumptions, we proposed three alternative forecasting models and we can choose the fittest based on the underlying relationship of linearity and non-linearity. The modeling process can be divided into two stage: In the first stage, SARIMA is mainly applied to fit the linear component. In the second stage, ANN is mainly employed to fit the nonlinear component and the left linear component in the residuals of linear forecasting values and actual values. Meanwhile, according to the autocorrelation in SARIMA, we construct the inputs of ANNs to improve the accuracy of forecasting. The monthly container throughput in Shanghai port(China) is applied into these models.Finally, we compare the forecasting performances of the seven models.The results indicate the hybrid 5 model has the best performance. Meanwhile, some theories are proved by the comparison of each models.