Abstract:
Accurate forecasting of port cargo throughput is critical for efficient logistics and strategic planning in global trade networks. This thesis presents an integrative study that employs multiple state-of-the-art forecasting methods both in univariate and multivariate settings to forecast total cargo volume at the Chittagong Port of Bangladesh. The methods include Long Short-Term Memory (LSTM) networks, Vector Autoregression (VAR), Seasonal Autoregressive Integrated Moving Average (SARIMA), and Transformer-based models to predict total cargo volume at Chittagong Port, Bangladesh. In addition, the research incorporates key economic indicator which is Gross Domestic Product (GDP) of Bangladesh, Export Cargo Volume and Import Cargo Volume as variables in a multivariate forecasting framework. An innovative component of this study is the incorporation of chaos theory to analyze the intrinsic nonlinear dynamics and sensitivity to initial conditions of the cargo throughput time series.
To uncover the underlying data dynamics, chaos theory is applied. By calculating Lyapunov exponents, Hurst Exponent and Entropy analysis, this thesis characterizes the degree of chaos, randomness, and complexity within the cargo throughput series. These non-linear analytical methods also provide an insight into the time frame of our data in terms of forecasting ability. This analysis provides critical insights into the stability and predictability of the system, further guiding model selection and parameter tuning. Moreover, the presence of chaotic behavior justifies the use of deep learning models like LSTM and Transformer, which are more tailored to capturing nonlinearity within the time series compared to traditional ARIMA methods.
This research commenced with a thorough exploratory dive into the monthly cargo statistics from Chittagong Port. Initial assessments, drawing upon foundational concepts of chaos theory, indicated that the cargo volume time series is not merely linear but instead displays moderately intricate dynamics. These dynamics are characterized by seasonal fluctuations, shifts and trends, and behaviors strongly suggestive of chaotic processes. Given these inherent complexities, the data clearly call for more sophisticated modeling approaches, as conventional linear methods would likely prove inadequate for capturing such nuanced patterns. To address these challenges, we proposed a hybrid forecasting method incorporating a sophisticated signal processing technique named Discrete Wavelet Transformer (DWT) and two deep learning models- Long Short Term Memory (LSTM) networks and another one is Transformer Network. Multiple settings were used in this research for forecasting. A univariate forecasting framework using individual models (ARIMA, LSTM) to predict cargo volume solely based on its historical values. Concurrently, a multivariate approach is implemented using the VAR model, LSTM model with external economic inputs (GDP, export, and import data). Then finally the hybrid model is introduced to the data for univariate settings. The focus of this thesis is to identify the adaptability of a hybrid forecasting method which DWT-LSTM and DWT-Transformer method, commonly known as Wavenet, which is the most recent analytical method invented. Also, this dual strategy allows for an assessment of model performance in isolated versus integrated settings with various other forecasting models. Moreover, all the models also help identify the additional predictive power of economic indicators.
In evaluating the performance of the various models, several metrics are employed, including Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE). These criteria not only quantify forecast accuracy but also facilitate the comparative analysis between univariate and multivariate configurations. The results show that the forecasting accuracy is the highest for DWT-LSTM (a MAPE value of 5.41%) followed by multivariate LSTM (a MAPE value of 6.32%), DWT- Transformer (a MAPE value of 7.01%), ARIMA (8.88%), univariate LSTM (11.19%) and finally VAR (11.97%). The results consistently through other metrices show that, while traditional ARIMA models provide a baseline level of accuracy, advanced models, particularly the hybrid LSTM approach, offer superior performance over all other models in capturing both long-term trends and short-term fluctuations.
The findings from this research offer some meaningful takeaways for port authorities and logistics planners. When forecasting is more accurate, it becomes easier to manage resources, schedule operations, and respond to demand shifts without unnecessary delays or bottlenecks. One particularly interesting insight is the presence of chaotic patterns in the cargo data. This suggests that the system is not just noisy or random rather it may actually follow a complex but deterministic path, which means sudden shifts could happen even when things seem stable. That kind of behavior makes a strong case for using adaptive strategies in port management, especially in environments as unpredictable as global trade. From an academic perspective, this study shows how blending machine learning techniques with tools from chaos theory can open up new ways of modeling complex logistics systems. Comparing univariate and multivariate models revealed that hybrid setups especially when paired with wavelet decomposition can capture more nuance and deliver better forecasts. Adding macroeconomic indicators like GDP, import, and export values added another layer of depth, helping to explain some of the broader trends behind cargo movements. And the chaos analysis, though more abstract, offered a deeper look into the structural behavior of the system over time.
In short, this thesis tries to bridge the gap between technical modeling and real-world application. It not only contributes to ongoing research in forecasting methods but also provides actionable insights that could genuinely help improve how ports operate in uncertain and dynamic conditions.