Data-driven multi-step solar photovoltaic predictions with limited and uncertain information: Insights from a collaboratively-optimized nonlinear grey Bernoulli model

Abstract

As solar energy grows as a crucial substitute for fossil energy worldwide, it is worth investigating solar photovoltaic predictions characterized by complex patterns and a lack of prior knowledge. Accordingly, an optimized nonlinear grey Bernoulli model is proposed, encompassing three essential advancements. Firstly, the damping accumulated generating operation is initially integrated with the grey Bernoulli model to reduce random disturbances and improve the smoothness of the original collections, extracting more valuable information via preprocessing the raw data. Secondly, the background value is modified to remove the jumping errors arise from converting differential to difference functions based on the Simpson rule. Thirdly, the heuristic intelligent algorithm is employed to collaboratively explore the optimal damping coefficient and power index in the proposed model. For illustration purposes, experiments on predicting the global photovoltaic installed capacity (case one) and solar electricity net generation (case two) are carried out to examine the effectiveness and generality of this new approach. Experimental results indicate that the proposed model achieves significant improvements over the prevalent models in one-to-three-step ahead forecasting, obtaining the average 323.94% and 211.86% improvement rates over the seven benchmarks measured by MAPE and RMSE. Furthermore, robustness tests in terms of Monte-Carlo simulation, algorithm comparison, and hyper-parameter sensitivity analysis are performed to explore the proposed model’s forecasting stability against the intelligent algorithm randomness, choices, and hyper-parameters. Generally, the proposed model is proven to be a reliable method and is deployed for future global solar photovoltaic predictions.