Short-Term Power Load Forecasting Under Unstable Data Quality: A Fuzzy Deep Neural Network With LSTM and Self-Attention

This article presents a deep neural network model that integrates long short-term memory (LSTM), fuzzy logic, and self-attention mechanism to enhance short-term power load forecasting, particularly in scenarios with unstable data quality. The architecture of the model is designed to harness the LSTM’s proficiency in sequential data analysis, the inherent fault tolerance of fuzzy logic, and the self-attention mechanism’s ability to selectively enhance the significance of pivotal information. In addition, a postcompensation mechanism is incorporated to compensate for the deviation when certain model’s parameters are unavailable. Validated on three datasets, the model demonstrates improved accuracy and superior robustness, even under poor data quality conditions.