Prediction of Distributed Photovoltaic Power Based on Feature Construction with TCN-Transformer

Accurate power prediction for distributed photovoltaic (PV) can improve power supply efficiency, optimize energy management and promote clean energy development. Due to the lack of meteorological measurement equipment for distributed PV, the prediction accuracy is limited. In this regard, this paper puts forward an approach that combines feature construction and model combination. Firstly, historical data are classified into three typical weather types using the K-means++ clustering algorithm; subsequently, temporal feature and state label are constructed through feature engineering to enrich the input information; finally, the TCN-Transformer model is introduced, which combines the temporal convolutional neural network (TCN) with the Transformer model to accurately predict PV power through dilated causal convolution and the self-attention mechanism. Simulation outcomes demonstrate that the approach presented is capable of effectively reducing the prediction error under different weather types without measured meteorological data.