A Short-Term Regional Net Load Prediction Method Based on Parallel Fragment Attention-BI-LSTM

With the increasing proportion of new energy sources on the user side such as distributed photovoltaics being connected to the grid, the difficulty in predicting the net load of the distribution network continues to intensify, making it difficult to ensure the source load balance and safe and stable operation of the grid. Therefore, a regional short-term net load prediction neural network model based on dual periodic attention mechanism and bidirectional short-term and short-term memory network is proposed to address the dual periodic fluctuations in photovoltaic power and user load implied in the net load. Firstly, a periodic attention neural network model with temporal segment similarity as attention weight was established. Then, the two types of information are independently encoded through a parallel attention mechanism to extract the periodic characteristics of photovoltaic power and user load fluctuations. Finally, the residual connection mechanism and Bi-LSTM are used for temporal decoding to obtain the net load prediction results. Case analysis shows that the proposed method has an average improvement of 1.84% in prediction accuracy compared to traditional methods.