Noise-Robust Distributed PV Power Prediction with Sampling-Noise Awareness

This paper presents a forecasting strategy for distributed photovoltaic (PV) systems using the echo state network (ESN), aiming to improve forecasting performance with data noise at sampling points of dispersed PV data collection devices. In real-world scenarios, noisy and non-normally distributed data caused by sensor malfunctions, human errors, and extreme weather lead to outliers, thereby undermining the reliability of power generation forecasting algorithms. Therefore, we propose an outlier-robust echo state network (OR-ESN) that leverages the combined strengths of ridge regularization and least absolute shrinkage and selection operator (LASSO) regularization to achieve enhanced generalizability and sparsity. Furthermore, by adopting the ℓ1-norm as the loss function, the model achieves robust performance in the presence of noise. In addition, as distributed PV systems are increasingly preferred because of their flexibility, efficiency, and economic advantages, we extend the OR-ESN to the distributed outlier-robust echo state network (DOR-ESN) to better meet the forecasting needs of distributed PV systems. Furthermore, the distributed average consensus protocol is introduced and combined with the ADMM algorithm to enhance communication efficiency during the training process of the distributed ESN network for power prediction. The experimental results show that the OR-ESN and DOR-ESN demonstrate excellent predictive performance. Specifically, the OR-ESN increases the resistance to perturbations due to noise in distributed PV power forecasting. Moreover, the DOR-ESN builds on this foundation to significantly improve the accuracy and adaptability of power predictions for distributed PV systems.