Hourly forecasting of solar irradiance based on CEEMDAN and multi-strategy CNN-LSTM neural networks

Abstract

Accurate and reliable solar irradiance forecasting can bring significant benefits for managing electricity generation and distributing modern smart grid. However, the characteristics of instability, intermittence, and randomness make an accurate prediction of solar irradiance very difficult. To exploit fully solar irradiance by the successful scheduling of electricity generation and smart grid, this work proposes a new CEEMDAN–CNN–LSTM model for hourly irradiance forecasting. Firstly, complete ensemble empirical mode decomposition adaptive noise (CEEMDAN) is employed to decompose original historical data into a set of constitutive series to extract data features. Secondly, a deep learning network based on convolutional neural network (CNN) and long short-term memory network (LSTM) is used to forecast solar irradiance in the next hour. Moreover, in this paper, the various CNN-LSTM-based strategies for solar irradiance forecasts are systemically investigated. Four real-world datasets on different climate types are employed to evaluate the full potential of the proposed model. Multiple comparative experiments show that the proposed CEEMDAN–CNN–LSTM model can accurately forecast the solar irradiance and outperform a large number of alternative methods. An average RMSE of 38.49 W/m 2 indicates that CEEMDAN–CNN–LSTM model has a relatively stable prediction performance in different climatic conditions. • A novel deep learning-based model is used to forecast multi-region solar irradiance. • CEEMDAN is used to decompose the original data in different climate regions. • Five types of CNN-LSTM are designed for multiple feature inputs. • The proposed model shows better accuracy than other state-of-the-art models.