Abstract
Seaport cranes operate continuously and consume large amounts of energy while aiming to minimise containerships' berthing time. Although previous studies have contributed to addressing the crane scheduling problem, most have focused exclusively on loading time, often overlooking the aspect of energy consumption. Furthermore, crane activity is typically modelled in a simplified manner-commonly assuming a fixed cycle duration or constant energy usage when handling a container-without accounting for the impact of variable container masses. In this study, an energy-aware quay crane scheduling formulation for container terminals is proposed, highlighting the importance of integrating an energy model into the scheduling problem. The optimisation problem is formulated as a Mixed Integer Linear Programming (MILP) model. The objective is to minimise total energy costs by reordering the sequence in which containers are handled, while respecting precedence constraints defined by the ship's stowage plan. Two solution methods-a MILP approach solved using CPLEX and a genetic algorithm (GA)-are compared. The results indicate that, for larger containerships, the genetic algorithm provides a more efficient solution method. Moreover, incorporating detailed energy consumption models for electric cranes may significantly reduce energy costs during containership handling operations.

Abstract

Abstract

Abstract
This project was developed to improve the cost estimation process of new products within the Product Development Department of a furniture manufacturer. This work involved developing a methodology using Machine Learning (ML) models trained on products' existing data to predict the cost of new innovative ones based on similarities and given data. The ML models used were Linear Regression (LR), Light Gradient-Boosting Machine (LGBM), Random Forest (RF), and Support Vector Machine (SVM). The proposed methodology considers the estimation of the total cost of producing a product, which encompasses both material and operational costs. Throughout this project, several analyses were developed to identify and evaluate different independent variables that could explain the behaviour of these two cost components. The suitability of the different variables was studied by applying several ML models, and a set of functions that return an estimate of the cost as a function of these predictor variables was obtained. The proposed approach, which incorporates ML models into more complex variables to predict, resulted in a 19.29% reduction in estimation error.

Abstract

Abstract
Non-Intrusive Load Monitoring (NILM) aims to estimate appliance-level consumption from aggregate household power measurements without requiring intrusive sensing infrastructure. However, most existing NILM approaches rely on offline supervisedlearning and static inference pipelines, which often struggle to adapt to changing appliance behavior, uncertainty in aggregatesignals, and domain shifts across households. This paper proposes an interactive NILM framework that integrates probabilisticappliance-level classifiers with a deterministic constraint-based inference engine and adaptive user feedback acquisition strategies.The proposed system operates in an online setting and selectively queries users about appliance states only when predictions areuncertain. To optimize this interaction process, we introduce a cost-sensitive active querying strategy based on a contextual bandit(LinUCB) that dynamically balances the expected improvement in inference accuracy against the cost of user queries. The resultsdemonstrate that cost-sensitive active querying provides an effective mechanism for improving NILM robustness under uncertaintyand evolving operating conditions. We use the REFIT dataset as the primary benchmark for real-world validation, and additionallyevaluate the framework on a synthetic household dataset generated with the open-source RAMP library. Across both datasets, theproposed strategy consistently reduces the Normalised Error in Power and Mean Absolute Error relative to a passive inferencebaseline– by 11–50% on REFIT, depending on household, and by 30.2% and 28.0% respectively on RAMP, alongside a 61.7%reduction in Signal Aggregate Error– while also outperforming a non-adaptive heuristic strategy using an identical query budget