A load prediction method using memory neural network and curve shape correction

Aiming at the problems that multiplex influencing factors and strong uncertainty in distribution network load caused by the capacity accumulation of distributed generation and new loads, a load prediction method using memory neural network and curve shape correction is proposed. In load peak prediction, the maximum information coefficient is applied to calculate the nonlinear correlation between load peak and influencing factors, so as to select the input features. Considering the long-term and short-term autocorrelation in load peak sequence and the different correlation between input features and load peak, the load peak prediction model is established with the Attention mechanism and bidirectional long-short term memory (BiLSTM) neural network. In load per-unit curve prediction, a prediction model is established by combining similar day and adjacent day through the reciprocal error method. In view of the non-stationary characteristics of prediction deviation, the complete ensemble empirical mode decomposition with adaptive noise and BiLSTM network are used to establish an error prediction model to correct the curve shape. The validity of the proposed model is verified by an example of regional power grid load of a city in northern China.