A Power Grid Service Activity Forecasting Method Based on GRU-PoolTransformer

With the ongoing advancements in electric power communication networks, the demand for constructing a more secure, stable, and reliable communication network is increasingly growing among State Grid Corporation of China, major power generation groups, and related tertiary industry enterprises to ensure the steady operation of electric power companies. Monitoring and predicting Power Grid Communication Networks Service Activity (PCNSA) holds significant reference value for identifying electricity usage anomalies, fault localization, preventing network congestion, and optimizing resource allocation. The inherent complex nonlinear relationships and significant short-term fluctuations in electric power time series data pose challenges for traditional single machine learning algorithms in predicting the trends of PCNSA changes. To overcome this challenge, this study conducted an in-depth analysis of electric power network traffic data and proposed an innovative GRU-PoolTransformer algorithm model for short-term PCNSA prediction. The model integrates the Transformer architecture, Gated Recurrent Unit (GRU) networks, and attention mechanisms, using bytes, packets, bits, access counts, sessions, and service activities as input features, extracting features through GRU and capturing the periodic characteristics and long-term dependencies of time series data using PoolTransformer networks. Experimental results demonstrate that the GRU-PoolTransformer hybrid deep learning model exhibits higher prediction accuracy in short-term forecasting tasks of PCNSA time series data.