Abstract:Due to the relatively limited research on the characteristics and prediction of solar radiation changes, and the fact that solar radiation has the most direct impact on photovoltaic power generation, studying the temporal changes and predictions of solar radiation in small areas lacks certain guiding significance for accurately estimating photovoltaic power generation. This article uses net solar radiation data from 1950 to 2020 in the vicinity of Tongxiang Photovoltaic Power Station. The minimum two component, M-K mutation, EEMD period decomposition, and BP neural network models are used to analyze the overall and interdecadal temporal variation characteristics and periodic evolution patterns, and to compare the applicability of various prediction models for predicting net solar radiation in this area. Research has shown that the overall net solar radiation shows a decreasing trend, with a decrease of 12.125MJ per square meter every 10 years. The most significant increasing trend was observed in the 1980s and 2000s, with 1956 and 2011 being years of sudden changes in net solar radiation; The net solar radiation exhibits multiple time scale periodic variations, namely 2-3 years (short-term), 10 years (medium-term), and 50 years (long-term); The BP neural network model established has a higher prediction accuracy compared to algorithms such as LSTM and DNN, and can be well applied in predicting the net solar radiation of the photovoltaic power plant. The research conclusion of this article can better reveal the evolution law of net solar radiation in small areas, which is of great significance for accurate calculation and prediction of photovoltaic power generation in photovoltaic power stations.