Construction and time series prediction of electricity bill accounting system combined with CNN and LSTM joint model

Abstract

With the advancement of smart grid and electricity market reform, electricity billing systems need to accurately integrate multi-source heterogeneous data (such as power load, meteorological information and user behavior) to achieve dynamic billing. However, existing models are difficult to balance between spatiotemporal feature extraction and dynamic load fluctuation capture, resulting in prediction results that are susceptible to data heterogeneity interference and delayed response to sudden load changes (such as extreme weather or holiday electricity consumption surges), affecting the realtime and fairness of electricity billing. To address the above problems, this paper proposes an improved CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory) joint model, which achieves high-precision time series prediction and robustness improvement through a multi-source heterogeneous data preprocessing module, a dynamic spatiotemporal feature fusion mechanism and a nonlinear load fluctuation prediction layer. Experimental results show that the improved model outperforms traditional ResNet (Residual Network), Transformer and basic CNN-LSTM models in terms of MAE (Mean Absolute Error) (0.85), RMSE (Root Mean Square Error) (1.10) and R² (0.92), and has low latency (82 ms) and high generalization ability. The median MAPE is as low as 4.0%-4.5% in the cold zone scenario of industrial users. This study provides a technical solution for the electricity billing system that takes into account both prediction accuracy and efficiency.