Time-series Electric Power Charge Load Prediction based on Graph Feature Fusion

In this paper, we introduce CBAT, a novel method for temporal electric charge load prediction that integrates bidirectional LSTM and a graph attention mechanism. The bidirectional LSTM module captures global sequence information, while the graph attention module, leveraging time steps as graph nodes, models the intricate relationships between these steps by constructing an attenuation factor adjacency matrix. This approach effectively captures local temporal relationships. Experimental results demonstrate that the CBAT model outperforms existing mainstream models in electric charge load forecasting and other time series prediction tasks, showing superior performance in RMSE, MAPE, and R2 metrics. Additionally, ablation studies confirm the efficacy of the time-series-based graph attention module in electric charge load prediction.