Power load forecasting based on LSTM and Attention Mechanism

Abstract In the context of promoting the “dual carbon” goals and constructing a new power system, power load forecasting is crucial for grid planning and new energy consumption. However, traditional methods mainly rely on historical load data, making it difficult to effectively integrate multi-dimensional dynamic factors such as meteorology and economy, resulting in problems such as insufficient prediction accuracy and poor noise resistance in scenarios with an increasing proportion of new energy. To address the above issues, this paper proposes a load forecasting model based on the combination of LSTM neural network and attention mechanism. The model captures temporal features through LSTM and dynamically weights key information by combining the attention mechanism, which significantly improves the fusion capability of multi-source heterogeneous data. The experimental results show that the model can maintain stable performance (R 2 ≈0.9) on different scale datasets such as industrial and commercial, and the prediction error is reduced by more than 20% compared with traditional methods. This study provides an effective solution for load forecasting under high proportion of new energy access, and its robustness and generalization ability can provide support for smart grid scheduling decisions. Future research will explore the adaptive optimization of the model in complex scenarios such as extreme weather.