Ultra-Short-term Load Forecasting Using TCN-LSTM with Photovoltaic Penetration Rate and Temporal Feature Extraction

To enhance the accuracy of ultra-short-term load forecasting while accounting for the impact of distributed energy resources, this paper presents a novel TCN-LSTM cascaded model that integrates photovoltaic penetration rates and multi-level temporal feature extraction for distribution transformer area load prediction. The proposed methodology first establishes a relationship model between PV generation and load consumption to quantify the influence of distributed PV systems on local power demand. Subsequently, a multi-level temporal feature extraction framework is introduced to capture multi-scale temporal patterns within load data. The correlation between load and temporal features is quantified using Mutual Information, with strongly correlated features selected as inputs for the Temporal Convolutional Network model. The architecture culminates in a TCN-LSTM cascade model, where TCN predictions and load data serve as inputs to the Long Short-Term Memory network. This design leverages TCN’s advantages in parallel processing and multi-scale feature extraction, while harnessing LSTM’s capability in capturing long-term dependencies. The empirical validation is conducted using load data from a distribution network in a city in southern China. Comparative analyses against standalone machine learning models demonstrate that the proposed method achieves superior load forecasting accuracy, robust stability, and enhanced dynamic response capabilities.