Imprint 2025

Hyperparameter Optimization of Decision Trees for  Solar Insolation Classification

G Ramesh, P Anbarasu, K Raj Thilak and Abhisek Paul

A proper classification of solar insolation is critical in order to maximize grid integration and operational planning in the fast growing solar energy market in India. Although decision-tree models, with their likely lauded interpretability, have significant opportunities when it comes to this task, their behaviour is critically dependent on sensible hyper - parameter tuning, which is currently under-investigated when it comes to solar classification. This paper aims to fill this gap by providing a systematic method of optimizing Decision-Tree classifiers to define solar insolation in India, exploiting an exhaustive dataset of Indian Climate and Energy Dashboard (ICED). Our optimization pipeline was strict and involved three stages, which included sensitivity analysis, methodological comparison, and extensive hyper-parameter optimization. In this process we tested 8320 unique model configurations by five-fold group cross validation. We found a better architecture, which uses an entropy criterion, allows unrestricted tree depth, and places only a small number of splitting restrictions (min_samples_split=2, min_samples_leaf=1), and that highly restricts the sampling of the features (max_features=0.7). This setting had a cross-validated macro-F1 score of 0.7130, the highest. Comparative analysis also determined Bayesian Optimization as the most efficient methodology, compared to exhaustive Grid Search by capturing 99.4%, and minimizing cost of calculation by 94%. Together, these findings highlight the importance of complex, unpruned trees in modeling the complex spatial-temporal variations of solar irradiation across the diverse geography of India and that our validated, high-performance and interpretable Decision-Tree model provides actionable parameter advice to improve the credibility of solar energy assessment and planning.

Paper Poster Presentation

AI-Driven Prediction of Water Potability for  Sustainable Human Consumption

B Vidya, M Nalina, B Brahma and Venkata Phanikrishna

Water is a vital resource essential for all forms of life, and ensuring its quality is crucial for public health and environmental sustainability. Traditionally, water quality assessment relies on time-consuming laboratory analysis of various parameters such as pH, hardness, total dissolved solids (TDS), chloramines, sulfates, and conductivity. These parameters often vary depending on the source - whether it's a river, canal, or treated water from storage tanks. Advancements in sensor technology now enable real-time collection of these parameters directly from water sources under varying environmental conditions. However, interpreting this sensor data to assess water potability still depends heavily on human experts and laboratory processes, which may lead to delays and inconsistencies. Moreover, water quality standards fluctuate over time due to environmental factors, necessitating continuous or periodic monitoring. To address these challenges and promote sustainable water resource management, this proposal explores the use of Artificial Intelligence (AI), specifically Machine Learning (ML), to automate and enhance water quality analysis. The proposed approach includes regression models to predict the Water Quality Index (WQI), classification models for determining Water Quality Class (WQC). The work also attempts to highlight the importance of Leakage Severity Index (LSI) which can be an imperative parameter in a scenario when only water quality data is available . By reducing dependence on manual analysis and enabling timely, data-driven decisions, this system aims to improve accuracy, efficiency, and long-term sustainability of water consumption. 

Key words: Water, Potability, Machine Learning, Water Quality Index (WQI), Water Quality Class (WQC), Classification, Regression

Paper Poster Presentation

Hybrid CNN–Random Forest Model for Efficient Gesture Recognition in AI-Driven Assistive Communication Systems

C Pratheeba, V Ravichandra and Deepika Gupta

Sign language represents the primary communication channel for people suffering from speech or hearing impairments. However, the differences in illumination, hand pose, and variations among users make the task of sign movement understanding by machines hard to realize. Traditional deep learning-based models, mainly Convolutional Neural Networks, have achieved good performance in constrained environments but suffer from a lack of generalization and computational efficiency when applied in unconstrained scenarios. In order to enhance the efficiency and accuracy of gesture recognition, this work presents a hybrid CNN-random forest framework. While the Random Forest classifier performs reliable decision-level classification, the CNN module is responsible for extracting high-level spatial features from American Sign Language gesture images. Experimental results demonstrate an accuracy of 97.4%, improving the baseline CNN models by approximately 5-7%, while reducing the inference time by 20%. The proposed framework can be easily integrated into real-time assistive communication systems and can further help in laying the foundation for multimodal AI-based solutions at the intersection of speech, vision, and AR.              

Keywords—Gesture Recognition, CNN, Random Forest, Hybrid Model, Sign Language, Assistive Communication, Deep Learning

Paper Poster Presentation

Analysis and Comparison of Machine Learning Methods for Crime Hotspot Detection

Samit Kumar Singh, Jagadish Patil, Amol Ahire and Deepika Gupta

Crime analysis studies criminal patterns to support law-enforcement decisions. As digitalization generates massive crime datasets, manual analysis becomes inefficient, increasing the need for automated, data-driven methods. This study proposes a machine-learning framework for Crime Hotspot Detection using the Vancouver Crime Dataset (2003–2017). Four models—Random Forest, KNN, Decision Tree, and Naïve Bayes—were applied to spatial and temporal features to identify high-risk zones. Random Forest performed best, showing machine learning’s usefulness for proactive policing and public-safety improvement.

Keywords— Crime Analysis, Machine Learning, Prediction, Data Mining.

Paper Poster Presentation

Thermal Radiation Bias Correction for Infrared Images Using Huber Function-Based Loss

Dhaval Bhojani, Naresh Vaghela and Jignesh Bhatt

Limited by the imaging mechanism, thermal radiation emitted from infrared (IR) imaging devices commonly contaminates the detector response to the scene, causing an additive thermal radiation bias at each pixel that dramatically reduces the contrast of the image. This degradation is considered a bias field over the image, impacting the visual perception and subsequent applications. Therefore, eliminating the thermal radiation bias field is an urgent issue. This article proposes an adaptive thermal radiation bias field correction method using Huber function-based loss, which can adapt to image contents and thus preserve meaningful details while eliminating the bias field. The proposed method introduces the low-order bivariate polynomial surface model to fit the bias field from the observed image precisely. We establish a robust objective function-based on the Huber function to estimate parameters of the surface model, which can adaptively switch between the ℓ1-norm and the ℓ2-norm-based loss functions according to the image region. Thus, our method not only effectively maintains optimality in flat regions but also improves robustness in edge and texture regions. To balance efficiency and robustness, we propose an adaptive threshold that controls the behavior of the Huber function. For stable convergence, an improved gradient descent strategy is utilized to solve the Huber loss-based objective function with the two-direction fitting and an adjustable step. Both simulated and real experiments against classical and state-of-the-art methods demonstrate the superior performance of the proposed method in improving the contrast and preserving details.

Paper Poster Presentation

AI-Based Management of Power Operations and Optimal Pricing Structure Selection in Competitive Electricity Markets

M S Khan, Bharat Kharinar, Manoj Patil and Pramit Mazumdar

This paper explores the integration of artificial intelligence (AI) into power system operations and electricity market pricing structures. With increasing penetration of renewable energy and competitive markets, traditional operational methods face challenges in efficiency, reliability, and cost optimization. AI techniques—including machine learning, reinforcement learning, and optimization algorithms—are applied to demand forecasting, real-time dispatch, congestion management, and dynamic pricing. The study demonstrates significant improvements in operational efficiency, cost reduction, and market competitiveness. Simulation results using standard test systems highlight the potential of AI-driven frameworks in achieving optimal generation scheduling and pricing strategies.

Paper Poster Presentation

Machine Learning Approach to Predict Carbide Banding and Interface Stability Using Process Parameters

Dharmeshkumar Tanti, Dharmesh Dudhatra and Jignesh Bhatt

Carbide banding and interface stability are critical metallurgical characteristics that strongly influence the performance, reliability, and service life of heat-treated steel components. Predicting these properties is challenging due to complex and non-linear relationships between heat treatment process parameters and microstructural behavior, while traditional trial-and-error methods are time-consuming and costly. This study proposes a machine learning–based approach to predict carbide banding and interface stability using key process parameters such as austenitization temperature, quenching temperature, chemical composition, holding time, and cooling rate. Exploratory data analysis confirms the suitability of data-driven modeling, and multiple machine learning techniques are employed to capture the underlying non-linear relationships, demonstrating the potential of machine learning for intelligent and data-driven metallurgical process optimization.

Paper Poster Presentation

AI based range estimation of an Electric Vehicle

Priyesh Pandey, Anurag Nijanandi and Pratik Shah

Electric vehicle (EV) range estimation remains a challenging problem due to the combined influence of environmental conditions, driving behavior, thermal loads, and battery operating states. Traditional physics-based models provide interpretable predictions but struggle to capture complex real-world effects, while machine-learning (ML) approaches risk overfitting and often report overly optimistic accuracy due to hidden data leakage across temporally correlated trips. This paper presents a Cross-Segment Leakage-Free (CSLF) hybrid framework that integrates a first-principles physics model with a data-driven residual learning module while explicitly preventing information leakage. The proposed method computes a physics baseline for aerodynamic, rolling, drivetrain, and auxiliary loads, and trains an ML model to predict the residual error between the baseline and the actual energy consumption. A GroupKFold segmentation scheme ensures that entire trips are isolated across training and testing, enabling robust generalization assessment. Experiments on a trip-level EV dataset demonstrate that the CSLF hybrid model significantly outperforms physics-only and ML-only baselines, achieving lower mean absolute error and more stable predictions on unseen routes. The results show that leakage-aware hybrid modeling provides a reliable and explainable solution for onboard EV range estimation.

Paper Poster Presentation