Short-term Photovoltaic Load Forecast Using RFE-CNN-BILSTM

Load forecasting guarantees the safe operation of photovoltaic systems and provides a basis for the effective management of electrical energy. It is needed to improve the accuracy of photovoltaic load forecasting by relying on historical load data as well as the influence of meteorology and radiance on it. In this paper, a short-term photovoltaic load forecasting method based on recursive feature elimination (RFE) and convolutional neural network (CNN) with Bi-directional long short-term memory (BILSTM) is proposed. Firstly, the factors with high importance are selected as the training data for load Firstly, the factors with high importance are selected as the training data for load forecasting, and then the CNN convolutional layer and pooling layer are used as the feature extract ion unit to extract the features that interact with each other in the data space. The reconstructed data is then input to the Bidirectional long short-term memory for feature mining. The convolutional neural network mechanism is introduced to extract high latent information of historical key moments autonomously with less number of weights, which test experiments based on load data from a region in eastern China show that the method has higher accuracy compared with traditional methods.