This paper studies a method combining long-short-term memory (LSTM) and decomposition-LSTM (DLSTM) networks with ensemble algorithms to predict solar flare occurrences using time series data from the GOES catalog. Using data from 2003 to 2023 (containing 151,071 flare events), we identify approximately 7,552 annual pattern windows, highlighting the difficulty of long-term prediction due to the complex and self-organizing criticality-based behavior of the Sun. A sliding window technique is used to detect temporal quasi-patterns in irregular and regular flare time series, and regularization reduces complexity, enhances large flare activity, and captures active days more effectively. A resampling technique is applied to address the class imbalance problem, and LSTM and DLSTM models are trained on peak flux and latency sequences of irregular time series, while LSTM and DLSTM combined with ensemble techniques are applied to sliding windows of regular time series with 3-hour intervals. The performance evaluation metrics (TSS 0.74, recall 0.95, area under the ROC curve AUC 0.87) demonstrate that DLSTM using ensemble techniques for regular time series outperforms other models, and provides more accurate large flare predictions with less error compared to models trained on irregular time series. The superior performance of DLSTM is attributed to its ability to effectively separate random noise by decomposing the time series into trend and seasonal components. This study highlights the potential of advanced machine learning techniques for solar flare prediction and emphasizes the importance of integrating different solar cycle phases and resampling strategies to enhance prediction reliability.
