Network load balancing method based on fuzzy clusterisation

The rapid expansion of interactive video, augmented and virtual reality platforms, and the Internet of Things sharply increases the requirements for flexible traffic distribution in modern networks. Conventional static load balancing algorithms fail to respond to rapid shifts in traffic patterns, leading to overloads on individual nodes. The study aimed to design and experimentally validate an adaptive balancing algorithm based on fuzzy grouping of server states. Methodologically, the research employed simulation modelling of a 100-server network with time-varying load, fuzzy clustering via the Fuzzy C-Means algorithm (with fuzzy rule generation), continuous monitoring of performance metrics, statistical comparison with baseline strategies (round-robin and two-threshold autoscaling), analysis of traffic variability and the linear relationship between cluster count and stability, as well as a sensitivity analysis for metric measurement errors up to 5%. It was found that the proposed algorithm reduces the mean deviation of node utilisation by a factor of 1.5 compared with two-threshold autoscaling. During peak periods, system response time decreased from 180 to 145 ms, while average resource utilisation rose from 68% to 85%. Analysis of the traffic coefficient of variation showed that at values above 0.4 the new method keeps node load within ± 10% for 92% of observations, whereas round-robin exceeds this range in 37% of cases. A linear relation between the number of clusters and distribution stability was revealed, with four clusters proving optimal. Fuzzy rules additionally eliminate abrupt traffic oscillations after demand spikes and ensure smooth flow redirection. Sensitivity analysis indicates that a metric measurement error up to 5% does not affect decision correctness. The resulting architecture maintains stable operation when individual servers experience disproportionate traffic growth. Comparison with traditional approaches confirmed the superiority of the proposed method across all evaluated performance and robustness metrics