Fully connected multi-reservoir echo state networks for wind power prediction

In this paper, considering the complex relationship between wind speed variation characteristics and data features, a fully connected multi-reservoir echo state network (FCMR-ESN) is proposed for wind power generation prediction, which can handle some issues such as insufficient extraction of data features and a gradual decline in memory capacity. Firstly, the Spearman correlation coefficient is used to calculate the correlation between the characteristic data. Secondly, the fully connected neurons are applied to the connections between the reservoirs. The fully connected layer can capture the complex nonlinear relationships in the wind power data and effectively map them to the prediction results. Thirdly, the reservoir parameters and the connection coefficients between the reservoirs are optimized using the improved weighted mean of vectors (INFO). Finally, The effectiveness of FCMR-ESN prediction is demonstrated through prediction experiments conducted on two sets of datasets with different lengths from different regions. In the one-day time prediction for the first set of data, the MAPE decreases to 2.90%. The MAPE for the three-day prediction in the second data set is reduced to 3.53%.