本研究針對時間序列提出多目標預測模型，結合卷積神經網路（Convolutional neural networks; CNN）與球型複數模糊神經系統（Sphere complex neural fuzzy system; SCNFS）。球型複數模糊集（Sphere complex fuzzy sets; SCFSs）可產生複數型態的歸屬程度，使SCNFS能根據目標數產生多個輸出值。在訓練資料進入模型前，使用多目標特徵選取，從中挑選有影響力之特徵。CNN置於SCNFS前，以容納更多有影響力之特徵，並識別特徵中趨勢的變化，萃取出有用資訊。參數訓練方面，使用高斯分布型鯨群最佳化演算法（Gaussian distribution based whale optimization algorithm; GD-WOA）及遞迴最小平方估計法（Recursive least squares estimator; RLSE）之複合式機器學習演算法，GD-WOA針對CNN權重參數及球型複數模糊集參數進行訓練；RLSE針對模糊神經系統之後鑑部參數進行訓練。為驗證本研究所提之方法，三個實驗以股價指數數據作為資料集，進行不同目標數的時間序列預測。實驗結果與過往文獻相比，有較佳的預測結果，顯示本研究所提之方法，在時間序列預測上有良好效能。

Purpose－Time series prediction is a challenging issue. Most prediction methods in literature implement prediction for a single target at a time only. In this paper, our purpose is to design a novel approach for multi-target simultaneous prediction. Design/methodology/approach－In this paper, the proposed approach integrate a convolutional neural networks (CNN) and a novel neural fuzzy system using sphere complex fuzzy sets (SCFSs). The proposed predictive model, denoted as CNN-SCNFS, is presented, where SCFSs can produce complex-valued membership degrees, providing more membership information than conventional fuzzy set can. The SCFS characteristics can enable the proposed model to bring about multi-output property for applications with multiple targets. A novel feature selection method is used for selecting useful cross-target data for the modeling of CNN-SCNFS. In the proposed CNN-SCNFS approach, a CNN is placed in front of the SCNFS, and the utility of CNN is to accommodate more useful features and to capture sudden changes of input data to the model. For machine learning, a hybrid method using the divide-and-conquer principle, called the GD-WOA-RLSE algorithm, is applied for parameter learning of the model. The Gaussian distribution based whale optimization algorithm (GD-WOA) hereby presented adapts the parameters of CNN and of SCFSs in the model while the well-known recursive least squares estimator (RLSE) updates the parameters of the Takagi-Sugeno consequence layer of the CNNSCFNS. Findings － Several real-world data sets of stock markets are used to test the proposed approach performing multi-target prediction. The results indicate that the proposed method has shown prominent performance through comparison with other methods. Research limitations/implications－In this study, the setting of proposed CNN architecture is determined by trial and error which may influent the performance of prediction. In the future work, we are intended to developed the model based on the input data to achieve self-organize property. Practical implications－Currently, data is growing faster than ever before. The data analysis become more important. If there are multiple targets we want to predict, we can use the proposed model to predict these targets simultaneously instead of using the model with single output property. Originality/value－In this paper, a novel approach for multi-target prediction is proposed which can well perform prediction for multiple targets in one model simultaneously.