Short-term Photovoltaic Power Prediction Based on Daily Feature Matrix and Deep Neural Network

Abstract

In order to reduce the error of short-term photovoltaic (PV) power forecast without irradiance data, a prediction model based on daily feature matrix and long short term-memory (LSTM) deep neural network is proposed. Firstly, various factors affecting PV output are analyzed to select model inputs effectively. On this basis, a new similar day selection method considering the internal and external factors under multi-source data integration scenarios is introduced. Based on weather forecast information and day-ahead PV power data, daily feature matrices can be constructed to determine similar days by calculating the distances between the matrices. Then, the similar historical PV power vector is used as an input of a LSTM deep neural network, combined with meteorological forecast information to realize the final power prediction. Finally, the feasibility of the proposed method can be validated with the actual data of residential PV systems in North America.