Short-Term PV Output Forecasting Based on Adaptive FCM Clustering and IKOA-CNN-BiLSTM

Introduction: The inherent uncertainty and random fluctuations of Photovoltaic (PV) power generation make high-precision output forecasting a critical technology for ensuring secure grid operation. To address this challenge, this paper presents targeted refinements to the key components of the PV forecasting pipeline. Methods: First, an adaptive factor is introduced to determine the optimal cluster number for the Fuzzy C-Means (FCM) algorithm. To enhance the rationality of cluster assignment, the standard Euclidean distance is replaced with a hybrid metric combining Mahalanobis distance and cosine angle distance. Based on the target prediction date, training and test sets are then allocated for each cluster. Subsequently, a hybrid prediction model is devised that integrates a Convolutional Neural Network (CNN), a Bidirectional Long Short-Term Memory (BiLSTM) network, and a Self- -Attention (SA) mechanism. For each cluster, this hybrid model is coupled with an Improved Kepler Optimization Algorithm (IKOA) to derive optimal hyperparameters during training. The final forecasts from all clusters are concatenated to generate the complete PV output curve. Results: Experimental results validate the effectiveness and superiority of the proposed hybrid method. Compared to traditional single-model and benchmark ensemble approaches, it achieves significantly higher prediction accuracy across all evaluated metrics. Discussion: Several limitations were identified during the research process: (1) lacking independent comparative experiments to validate the superiority of enhancement strategies specifically within the PV forecasting domain. (2) Certain experimental design aspects could be refined. (3) The increased complexity of the model architecture substantially extended the runtime. Conclusion: This study advances prior research by refining the FCM clustering method to achieve more coherent cluster formation, thereby improving downstream prediction accuracy. Furthermore, it proposes an improved meta-heuristic algorithm tailored for PV forecasting, addressing specific limitations of the original optimizer to provide superior hyperparameter configurations that maximize model performance. Finally, the case study results validate the prediction accuracy of the hybrid model, demonstrating the synergistic benefits of integrating the enhanced clustering algorithm with the proposed hyperparameter optimization technique.