Much research has focused on financial market time series. This thesis presents a comprehensive analysis of gold price prediction and market structure using a visibility graph network. With a specific focus on the years 2008 and 2024, we discuss the price dynamics of gold in two currencies: Bangladeshi Taka (BDT) and US Dollars (USD), using two complementary methods: Long Short-Term Memory (LSTM) forecasting and Visibility Graph (VG) analysis, which revealed the structural complexity of the market. The LSTM model captures the non-linear movements of highly volatile gold prices, giving it strong predictive capability and providing useful insights for investors and policymakers in Bangladesh. Golds popularity as a safe-haven investment during economic uncertainty underscores the importance of forecasting, and our results show that LSTM can generate effective long-term projections. At the same time, by constructing gold price time series as complex networks using the VG algorithm, we study structural characteristics every year. Important network measurements-Average Clustering Coefficient, Mean Degree, Modularity, Average Shortest Path Length, Structural Entropy and Global Reaching Centrality are calculated to find underlying structures and market dynamics. Comparative analysis between BDT and USD markets reveals that the Bangladeshi gold market exhibits higher modularity and entropy, indicating a more diversified structure. Both markets show signs of stability over time, with significant structural deviations during major global financial events. Degree distributions follow a power law in most years, indicating fractal characteristics in gold price dynamics. Together, the LSTM forecasting and visibility graph-based structural analysis offer a dual perspective on the gold market— capturing both future price trends and the underlying market complexity. These insights not only deepen our understanding of gold market behaviour but also provide a foundation for informed decision-making in investment and economic policy.
Climate change poses increasing risks to agricultural systems in semi-arid regions through rising temperatures, changing rainfall variability, increasing crop-water requirements and growing pressure on groundwater resources. This study assesses potential future climate-related risks to agricultural sustainability in Hisar district, Haryana, using a scenario-based framework that combines the historical climatic context of 1990– 2025 with published CMIP6-informed scenario information under SSP2-4.5 and SSP5-8.5. We consider representative future periods to examine how climate-related agricultural risks progress. We translate scenario-based temperature changes into indicative crop-productivity responses using previously established Hisar-specific crop-temperature elasticities for wheat, mustard, cotton, and bajra. The assessment further examines the potential implications of warming and rainfall variability for agricultural water requirements and groundwater stress. Results indicate that wheat exhibits the greatest temperature-related sensitivity among the crops considered, followed by mustard, while cotton and bajra show comparatively lower direct temperature sensitivity. The findings also indicate that rising temperatures may increase evapotranspiration and irrigation needs, potentially intensifying groundwater extraction pressure in areas already experiencing groundwater stress. The study highlights the interconnected nature of climate, crop productivity, and agricultural water security in Hisar. It emphasises the need for heat-resilient crop planning, efficient irrigation, groundwater conservation, and integrated climate–water management. The study provides a district-level scenariobased assessment that can support future climate-resilient agricultural planning and more detailed modelling using highresolution climate, crop-water and groundwater datasets.
In recent years, deep learning has become a fundamental technology across a wide array of scientific and industrial fields, largely fuelled by advances in computational capabilities. One area that has experienced substantial progress is face hallucination—the task of improving the resolution of facial images. This process is critical to various computer vision applications, including facial recognition, feature extraction, and identity verification. Recently, deep generative models, particularly Generative Adversarial Networks (GANs), have led the field. Although these models have produced remarkable results, there is still a pressing need to further improve both accuracy and output quality. In order to address these problems, we propose a new GAN-based face hallucination method. This method is primarily based on the Enhanced Super-Resolution Generative Adversarial Network (ESRGAN). We present a personalised adaptation of ESRGAN that employs the VGG16 architecture with a compact pre-trained version. This method balances output image quality and computational efficiency. Experiments show that our approach is effective. The improved model obtains a maximum peak signal-to-noise ratio (PSNR) of 30.30. The Learned Perceptual Image Patch Similarity (LPIPS) score is 0.0817, whereas the Structural Similarity Index Measure (SSIM) is 0.8757. The results surpass many state-of-the-art methods available today. These enhancements have a significant impact and importance.
This study presents a prototype that classifies full EEG recordings as either seizure-labelled or non-seizure-labelled. The method first uses Fast Fourier Transform (FFT) to convert each EEG into the frequency domain, then applies a two-dimensional Convolutional Neural Network (2D CNN) for classification. It was observed that many trainees and students found EEG interpretation difficult. Labelling was straightforward: any recording with at least one labelled seizure was marked as seizure labelled, and the rest as non-seizure-labelled. Training used two public datasets from PhysioNet: the pediatric CHB-MIT and adult Siena scalp EEG databases, with seizure labels from patient annotation files. To give the prototype a realistic run, eleven anonymised EEG recordings from Aster Royal Al Raffah Hospital in Muscat, Oman, were taken and used for external validation; out of those eleven, five were labelled as non-seizure recordings and six as seizure recordings. In early tests, the prototype achieved about 88% accuracy, with F1-scores of 0.95 and 0.91 for the two classes, and 82% accuracy on the hospital data. These results are preliminary, and full validation is planned. The tool is designed to help trainees during supervision, not to diagnose patients or identify seizures within recordings.
Overhead transmission line conductors are expected to maintain mechanical integrity and predictable thermal performance over multidecade service lives. Operation under harsh environmental conditions can accelerate material degradation and reduce the reliability of power delivery. This paper presents a laboratory investigation of material-driven failure mechanisms in field-aged overhead conductors extracted from long-serving transmission lines in Saudi Arabia. Multiple conductor designs, including AAAC, ACAR, ACSR (greased), ACSR/AW, and GAP, were evaluated after exposure to coastal, coastal-urban, desert-coastal, and urban environments. The study integrates layer-by-layer visual inspection, bending tests, galvanizing retention and cladding verification, torsion and tensile testing, and surface emissivity measurements. Results show that unprotected aluminium conductors exhibit severe outer-layer embrittlement and visible corrosion products, whereas greased and aluminium-clad designs retain ductility and protective coatings.
