Attention-Enhanced Network Traffic Prediction for Intelligent Load Balancing Optimization

Abstract

This paper focuses on the problem of load balancing optimization in complex network environments and proposes a time series-based network traffic prediction method to achieve more efficient resource scheduling and traffic management. The proposed model integrates gated recurrent units (GRU) with a multi-head self-attention mechanism to capture both short-term local fluctuations and long-term dependencies. In data preprocessing, the model applies a sliding window and normalization strategy to model multi-node traffic sequences, enhancing adaptability in dynamic scenarios. During training, a weighted mean squared error is used as the loss function to guide the model in accurately identifying high-variance regions, thereby improving overall prediction quality. To validate the effectiveness of the method, experiments are conducted on the MAWI network traffic dataset, including a sensitivity analysis on time window length and sampling frequency. Experimental results show that the proposed method outperforms mainstream models across multiple error metrics and demonstrates strong robustness under different parameter settings and data conditions. This work provides accurate and stable predictive support for intelligent load balancing and shows strong potential for deployment in real-world network scenarios.