Venkateshwaran Bakthavachalam

There are urgent problems in urban sustainability monitoring to process huge heterogeneous data streams of IoT sensors, building management systems, and environmental monitors and ensure analytical accuracy in real time. The proposed study presents Count-MinSketch with SHAP-Driven Adaptive Sketch Optimization (SDASO), which is a new probabilistic data structure that is augmented with explainable AI to support real-time smart city sustainability assessment. The suggested framework combines space-efficient approximation of Count-Min Sketch with SHAP (SHapley Additive exPlanations) feature attribution that dynamically chooses parameters of the sketch according to the features of data streams. When summarized using Count-Min Sketch, Heavy Hitters (Lossy Counting), the Sliding Window Algorithm, and ADWIN (Adaptive Windowing), it is seen that it performs better when compared to other algorithms in terms of computation and accuracy. The SDASO model had an outstanding approximation quality of 98% and a small error rate (δ=0.012) and approximation error of 0.003 with a root mean squared error of 1375.92. The SHAP integration produced impressive model explainability of R² = 0.9957 and an effectiveness score of 0.1223 and was capable of selecting adaptive features through 9 indicators of critical sustainability. Evaluation of the system performance indicated exceptional computational performance: throughput was 367.04 records/second, processing latency was 2.72 ms, CPU utilization was only 2%, and the total memory footprint was 0.095 MB operating with 2,673 building records and 8 partitions with a sketch utilization of 47.46%. The findings indicate that SDASO is able to deliver interpretable, resource-effective, and highly accurate real-time sustainability monitoring of large-scale smart city infrastructure to support the realization of data-driven urban environmental governance and policy optimization.

This study introduces a hybrid framework combining Betweenness Centrality (BC) and a Multi-Objective Evolutionary Algorithm based on Decomposition with Differential Evolution (MOEA/D-DE) for intelligent sustainable manufacturing monitoring and optimization. The presented model overcomes major issues in big data analytics of complex manufacturing networks by focusing on accuracy and performance of network analysis simultaneously. The conventional algorithms of BC have scaling issues when dealing with large manufacturing datasets; the hybrid architecture will eliminate those obstacles by using evolutionary multi-objective optimization. The model has outstanding performance with an approximation error at 0.012187 with a mean squared error at 0.0017, with a 3990.91% improvement in BC accuracy over baseline methods. The MOEA/D-DE has the best convergence (0.99) and optimized sampling rate (0.1549) and depth limit (2); the error rate is the lowest (0.0069). The efficiency of the approach is confirmed by the system-level measures: the CPU time of 0.0013 seconds, CPU utilization of 74 percent, processing time of 0.094 seconds, and ultra-low latency (16.87 ms). The platform has 10,119.46 records/second processing throughput, and memory consumption is relatively low (1255 MB). The model is highly fault tolerant, and its performance is equally scalable (8.2/10) with networks that have 110 nodes and 34 edges. These findings show that there is strong potential for real-time optimization of manufacturing networks to allow the use of data to make decisions that are sustainable in manufacturing. Combining evolutionary algorithms and centrality can present a powerful method of discovering the critical nodes and bottlenecks within the manufacturing systems to enable proactive sharing of resources and process optimization in Industry 4.0 systems.

The development of increased complexity of environmental problems facing the world demands the development of more advanced and powerful monitoring tools. The research meets that requirement by offering a new hybrid machine learning methodology and comparing different ensemble models in a systematic manner. One of them, Graph Neural Network + CatBoost (GNN + CatBoost), produced the most promising results. This model was then optimized with the optimization algorithm known as AdaGrad to come up with the final GCAO proposed model. The GCAO framework established new standards in performance with the maximum training accuracy of 98.53 percent and a high test accuracy of 90.91 percent. In addition to its great prediction accuracy, the GCAO model was also established to be statistically reliable. The benefit of AdaGrad component addition is found to be statistically significant, with the ANOVA test giving a p-value of 0.005. Such solid statistical data confirms the fact that the novel GNN+CatBoost with AdaGrad is not only the most accurate among the optimization variants that are compared to it but also better in terms of its reliability and the overall model stability. It, therefore, offers a very effective and proven means of handling complex international environmental monitoring issues.