ARIMA-2DLSTM identification method for key characteristics of distributed photovoltaic output based on similar days

In order to better solve the problems of low correlation, large variation, and nonlinear interference in the identification process of key features of distributed photovoltaic output, this paper proposes a recognition method that combines similar day selection and ARIMA-2DLSTM fusion architecture. Using the similar day strategy to select output data with high climate matching in previous examples has increased the relevance and effectiveness of input information. The ARIMA model is used to extract mid - to long-term trend features from time-series data, achieving stable linear modeling results. The 2D-LSTM architecture models data in terms of time and features respectively, making the extraction of local disturbances and nonlinear features more accurate. The fusion module receives the feature results of the trend and outputs them to form a joint model, which improves the recognition accuracy of the entire model. Compared with traditional methods, this method has significant improvements in indicators such as RMSE, MAPE, and R2, and has strong practical value.