Latency Prediction in Distributed Control Systems Using FPGAAccelerated Neural Networks

Abstract

To improve the real-time performance and stability of distributed control systems in complex and dynamic environments, this study introduces a delay prediction and optimization model. The model is built on an integrated architecture that combines Long Short-Term Memory (LSTM) neural networks with Field Programmable Gate Array (FPGA). A sliding window input mechanism is used, where a recent sequence of historical delay data serves as input to forecast short-term system response latency. To support efficient hardware deployment, the LSTM model was quantized to 8-bit fixed-point precision. Additionally, the FPGA implementation was optimized through the design of a parallel pipelined architecture and an onchip cache scheduling mechanism. These enhancements significantly improve inference speed and resource utilization. Experiments were conducted using the Electric Transformer Temperature (ETT) time-series dataset series. The proposed model was compared against several representative approaches. Evaluation metrics included prediction accuracy, response latency, system throughput, resource consumption, task success rate, and overall stability. On the ETT-small-m3 dataset, the optimized model achieved a task completion rate of 99.699%, a system throughput of 1,424.082 tasks per second, and an average response time of 0.247 seconds. These results surpassed those of the baseline models across most performance indicators. To evaluate generalization, five-fold cross-validation was performed. Analysis of variance (ANOVA) was also conducted to confirm the statistical significance of the results, with all pvalues below 0.05, ensuring the reliability of the experimental findings. Despite its strengths, the model has limitations in certain reliability metrics. For example, the mean time between failures was slightly lower than that of the Multi-Agent System-Based Distributed Control Model (MAS-DCM), suggesting reduced stability under high-pressure or high-load conditions. Moreover, the model's adaptability to scenarios involving multi-source heterogeneous data has not been comprehensively tested. In summary, this study presents a deployable, efficient, and scalable architecture for intelligent delay prediction. The proposed solution provides a practical approach to delay modeling and performance optimization in smart control systems. It holds strong potential for real-world applications and lays a solid foundation for future research and development in this area.