An integrated data-driven framework for predicting degradation trends of pumped storage units

Pumped storage units (PSUs) play a critical role in ensuring the stability and security of power grids. However, during long-term operation, the performance of PSUs may gradually degrade, thereby posing potential threats to the safe and stable operation of the power system. Therefore, accurate prediction of the degradation trend of PSUs is essential. Nevertheless, the degradation process exhibits significant nonlinear and non-stationary characteristics, which makes it challenging for a single prediction model to effectively capture its complex dynamic features. To address this challenge, this paper proposes a novel framework for degradation trend prediction of PSUs. Firstly, polynomial XGBoost is employed to construct the health state model, and the Shapley additive explanations (SHAP) method is applied to interpret the contribution of each input feature to the model output, thereby improving the model's interpretability. Next, based on the output of the health state model, a performance deterioration index (PDI) is constructed to characterize the degradation level of PSUs. Then, improved water uptake and transport in plants (IWUTP) algorithm incorporating elite selection, crossover and mutation is proposed to optimize the parameters of variational mode decomposition (VMD), thereby enabling efficient decomposition of the PDI. Following this, a two-layer predictor with stacked features is developed, in which the first layer utilizes gated recurrent unit (GRU) to extract temporal characteristics from the PDI sequence, while the second layer applies IWUTP-optimized kernel extreme learning machine (IWUTP-KELM) to perform ensemble learning on the GRU-derived features, ultimately yielding the degradation trend prediction results.