Lightweight Multi-Channel BiLSTM-Attention with Coordinated Defense for Intrusion Detection in Distributed Photovoltaic Communication Networks

Abstract

To address intrusion and data tampering threats in distributed photovoltaic (PV) communication networks, this paper proposes a lightweight multi-channel BiLSTM-Attention–based intrusion detection and coordinated defense framework. An attack-oriented, risk-aware feature modeling method constructs multi-dimensional temporal inputs by combining protocol semantics, operational interaction behaviors, and deviation-based risk indicators, which are processed by a BiLSTM backbone with time-step attention and channel-level attention fusion. Model pruning and attention sparsification are adopted to obtain a compact architecture suitable for edge deployment, while a dual-threshold, detection-driven defense coordination mechanism links anomaly scores to graded response policies and closed-loop control. Experiments on a hybrid dataset composed of real PV communication logs and PV-adapted UNSW-NB15 and CIC-IDS2018 traffic show that the proposed model achieves 94.8% accuracy, 94.2% recall, and a 95.6% F1-score with an average inference time of 16.2 ms and a false positive rate below 3.6%. Compared with a vanilla BiLSTM-Attention baseline, the proposed framework improves F1-score by 3.2 percentage points and reduces inference latency by about 11%, while maintaining higher robustness under noisy and imbalanced conditions. These results indicate that the framework can provide accurate, real-time intrusion detection and coordinated security defense for distributed photovoltaic communication networks in resource-constrained deployment scenarios.Povzetek:An improved lightweight multi-channel BiLSTM-Attention framework enhances intrusion detection and coordinated security defense in distributed photovoltaic communication networks through risk-aware temporal modeling, multi-channel attention fusion, and model pruning for edge deployment.